CN116029407A - Taxi travel demand prediction method based on ConvLSTM - Google Patents

Abstract

The invention discloses a taxi travel demand prediction method based on ConvLSTM, which comprises the following steps: acquiring historical traffic data of a region to be predicted and preprocessing the data; dividing the area to be predicted according to the longitude and latitude directions, acquiring unit historical traffic data of each unit interval, and carrying out normalization processing; acquiring a test set and a training set; inputting the space transverse and longitudinal analysis matrix into a taxi travel demand prediction model based on ConvLSTM; acquiring a loss function; acquiring accuracy between the test predicted traffic data and the real traffic data; and outputting the predicted traffic data within the future time threshold range. According to the invention, through carrying out standardization and normalization processing on the historical traffic data of each unit interval, the operation amount is reduced, and the prediction efficiency is improved. And the time sequence relationship can be established like an LSTM, so that the accuracy of taxi track prediction is greatly improved.

Description

A taxi travel demand prediction method based on ConvLSTM

Technical Field

The present invention relates to the field of artificial intelligence technology, and in particular to a taxi travel demand prediction method based on ConvLSTM.

Background Art

In recent years, with the rapid increase in the number of motor vehicles and the continuous growth in the number of taxis, the convenience of residents' travel has also been continuously improved. Taxis not only provide convenient "door-to-door" services for residents' travel, but are also an important supplement to conventional public transportation. However, the expansion of the existing road network is subject to many factors, and its scale does not have the possibility of unlimited expansion. According to the historical and current growth of the number of taxis, the number of taxis will not increase much, resulting in an increasingly intensified contradiction between residents' travel demand and taxi service supply. Therefore, it is very important to predict the future travel demand of taxis based on the historical travel demand data of taxis.

The urban intelligent transportation system is a multi-dimensional complex system that integrates computing, network and physical environment. As far as we know, the research methods of previous people for predicting taxi travel needs can be roughly divided into the following two types: One is to use traditional methods for prediction, such as the prediction method based on Markov chain, that is, to perform data statistics on historical data and calculate its transfer probability. The prediction accuracy of this method is unsatisfactory. In addition, when the traffic volume is small, this method cannot make predictions; the second type is to use neural networks for prediction, such as WU et al. used neural networks to give the same weight to each road section, so that they can be clustered for prediction. However, this method is not only based on the prediction of known road sections, but also the network model is too complex and has certain limitations. CHEN et al. used a prediction method based on the attention mechanism traffic volume prediction to add an attention mechanism to LSTM, but this method model is also relatively complex and lacks its transferability.

Summary of the invention

The present invention provides a taxi travel demand prediction method based on ConvLSTM to overcome the above technical problems.

A taxi travel demand prediction method based on ConvLSTM includes the following steps:

S1: Establish a taxi travel demand prediction model based on ConvLSTM:

S2. Obtain historical traffic volume data of the area to be predicted within the historical time range threshold, and pre-process the historical traffic volume data to obtain standard historical traffic volume data; the historical traffic volume data includes the taxi license plate number, passenger status, passenger time, longitude of the taxi location, latitude of the taxi location, and speed in kilometers per hour;

S3, dividing the area to be predicted according to the directions of longitude and latitude to obtain a number of unit intervals, and obtaining unit historical traffic volume data within the historical time range threshold of each unit interval according to the standard historical traffic volume data;

S4, normalizing the unit historical traffic volume data to obtain the normalized unit historical traffic volume data;

S5: Divide the area to be predicted into a spatial horizontal and vertical analysis matrix of n*n tuples, so that the normalized unit historical traffic volume data corresponds one-to-one to each tuple, the rows of the spatial horizontal and vertical analysis matrix represent the longitude of the location of the area to be predicted, the columns of the spatial horizontal and vertical analysis matrix represent the latitude of the location of the area to be predicted, and the elements in the spatial horizontal and vertical analysis matrix are the unit historical traffic data of the corresponding unit interval;

S6. Obtaining a test set and a training set according to the normalized unit historical traffic volume data;

S7: According to the training set, the spatial horizontal and vertical analysis matrix is input into the taxi travel demand prediction model based on ConvLSTM to obtain the trained taxi travel demand prediction model based on ConvLSTM;

S8: Input the test set into the trained ConvLSTM-based taxi travel demand prediction model to obtain test predicted traffic volume data; obtain the loss function between the test predicted traffic volume data and the real traffic volume data. When the value of the loss function is less than the loss function value threshold, execute S9. At this time, the trained ConvLSTM-based taxi travel demand prediction model is the optimal ConvLSTM-based taxi travel demand prediction model; otherwise, repeat S7;

S9: Obtain the accuracy between the test predicted traffic volume data and the actual traffic volume data; and obtain the predicted traffic volume data within the output future time threshold range according to the optimal ConvLSTM-based taxi travel demand prediction model.

Furthermore, in S4, the normalized unit historical traffic volume data is obtained as follows:

Where _yk is the normalized unit historical traffic volume data of the kth unit interval, _tk is the unit historical traffic volume data of the kth unit interval, max(t) is the maximum value of the unit historical traffic volume data, min(t) is the minimum value of the unit historical traffic volume data; k is the number of the unit interval.

Furthermore, the loss function between the test predicted traffic volume data and the actual traffic volume data in S8 is:

为基于ConvLSTM的出租车出行需求预测 模型的第i个预测值，output size为基于ConvLSTM的出租车出行需求预测模 型的参数个数；Where Loss is the loss function value; _yi is the i-th element value of the unit historical traffic volume data of the area to be predicted after normalization;

is the i-th predicted value of the taxi travel demand prediction model based on ConvLSTM, and output size is the number of parameters of the taxi travel demand prediction model based on ConvLSTM;

Furthermore, the accuracy between the test predicted traffic volume data and the actual traffic volume data in S9 is calculated as follows:

为基于ConvLSTM的出租车出行需求预测模型的第i个预测值； ABS(·)表示取绝对值运算。In the formula, _yi is the i-th element value of the unit historical traffic volume data of the area to be predicted after normalization;

is the i-th predicted value of the taxi travel demand forecasting model based on ConvLSTM; ABS(·) represents the absolute value operation.

Beneficial effects: The taxi travel demand prediction method based on ConvLSTM of the present invention reduces the amount of calculation and improves the prediction efficiency by standardizing and normalizing the historical traffic volume data of each unit interval. The ConvLSTM neural network used can not only establish a temporal relationship like LSTM, but also characterize local spatial features like CNN, which greatly improves the accuracy of taxi trajectory prediction.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

FIG1 is a flow chart of a method for predicting taxi travel demand according to the present invention;

FIG2 is a ConvLSTM model architecture diagram of the present invention;

FIG3 is a flow chart of a taxi travel demand prediction model based on ConvLSTM of the present invention;

FIG. 4 is a line graph of accuracy-iteration number in an embodiment of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

This embodiment provides a taxi travel demand prediction method based on ConvLSTM, as shown in FIG1 , comprising the following steps:

S1: Establish a taxi travel demand prediction model based on ConvLSTM:

Specifically, the formula of the taxi travel demand prediction model in this embodiment is as follows:

表示对应元素相乘；In the formula, _Wx is the weight parameter of the input gate of the ConvLSTM network; _Wh is the weight parameter of the output gate of the ConvLSTM network; _Wc is the weight parameter of the forget gate of the ConvLSTM network; _bi is the bias parameter of the input gate of the ConvLSTM network; _bf is the bias parameter of the output gate of the ConvLSTM network; _bo is the bias parameter of the forget gate of the ConvLSTM network; _xt is the input at the current moment; _it is the value of the input gate at the current time step; _ct is the value of the forget gate at the current time step; ct _-1 is the value of the forget gate at the previous time step; _ft is the value of the memory cell at the current time step; _ot is the value of the output gate at the current time step; Ht _-1 is the hidden state at the previous time step; σ is the sigmoid activation function; tanh is the trigonometric tangent function; * indicates convolution, which is different from the general LSTM;

Indicates the multiplication of corresponding elements;

Specifically, ConvLSTM, or Convolutional Long Short-Term Memory Network (CNN-LSTM Network), is a hybrid network model based on Convolutional Neural Network and Long Short-Term Memory Network system, where Convolutional Neural Network (CNN) is a type of feedforward neural network with deep structure that includes convolution calculation, and mainly realizes feature extraction; Long Short-Term Memory Network (LSTM) is a time recurrent neural network, which is specially designed to solve the long-term dependency problem existing in general RNN (recurrent neural network). All RNNs have a chain form of repeated neural network modules, and mainly realize the prediction of time series. In the ConvLSTM model, the input of the convolutional neural network part is the historical traffic data of taxis, the output of the convolutional neural network part is the feature map extracted by the convolution layer, the input of the Long Short-Term Memory Network part is the feature map output by the convolutional neural network part, and the output of the Long Short-Term Memory Network part is the predicted value of the future travel demand of taxis. The use of RNN to carry out the taxi demand forecasting method in this embodiment can effectively solve the problem of low forecasting accuracy caused by changes in the number of historical and existing taxis.

The core concepts of ConvLSTM are cell states and "gate" structures. Cell states are equivalent to the path of information transmission, allowing information to be passed down in the sequence, and can be regarded as the "memory" of the network. In theory, cell states can pass on relevant information in the sequence processing process. Therefore, even information from earlier time steps can be carried to cells in later time steps, which overcomes the influence of short-term memory. The addition and deletion of information is achieved through the "gate" structure. The "gate" structure will learn which information should be saved or forgotten during the training process. The cell structure of the ConvLSTM model is shown in Figure 2: Gate 1 is the forget gate, which is used to determine which relevant information in the previous step needs to be retained, Gate 2 is the input gate, which is used to determine which information in the current input is important and needs to be added, and Gate 3 is the output gate, which is used to determine what the next hidden state should be.

S2. Obtain historical traffic volume data of the area to be predicted within the historical time range threshold, and pre-process the historical traffic volume data to obtain standard historical traffic volume data; the historical traffic volume data includes the taxi license plate number, passenger status, passenger time, longitude of the taxi location, latitude of the taxi location, and speed in kilometers per hour;

Specifically, in this embodiment, the area to be predicted is Dalian City. The historical traffic volume data of taxis in the entire Dalian City within the historical time range threshold are collected by GPS installed on taxis, and the original historical traffic volume data of taxis are preprocessed. The preprocessing is to clean the missing values, abnormal values and duplicate values so that the historical traffic volume data can match the requirements of the taxi travel demand prediction model to adapt to the taxi travel demand prediction model. The specific preprocessing method includes coordinate screening, data cleaning and de-duplication operations, and then the standard historical traffic volume data of taxis in the area to be predicted are counted as the input of the subsequent model.

S3. Divide the area to be predicted according to the directions of longitude and latitude to obtain a plurality of unit intervals, and obtain the unit historical traffic volume data of each unit interval within the threshold value of the historical time range according to the standard historical traffic volume data; wherein, when the area to be predicted is divided according to the directions of longitude and latitude, the selected units of longitude and latitude are specifically determined manually according to the local terrain conditions of the selected area to be predicted.

Specifically, in this embodiment, the groupby function is used to count the unit historical traffic volume data of each unit interval; the unit historical traffic data in the unit interval can be normalized within the set historical time range threshold, and the elements in the 250*250 matrix corresponding to a certain moment within the historical time range threshold can be obtained, wherein the historical time range threshold is a set time period in the past. In full consideration of the terrain of Dalian, Dalian is divided according to the direction of longitude and latitude, and the range of 0.001 units of longitude and latitude is taken as a unit interval, and 5 minutes is taken as a time range threshold, that is, 5 minutes is taken as a granularity, and the unit historical traffic volume data of taxis in a unit interval (i.e., the area surrounded by 0.001 units of longitude and latitude) within every 5 minutes is counted, and then the unit historical traffic volume data of taxis in each unit interval of the entire Dalian city in every 5 minutes is counted.

S4. Normalize the unit historical traffic volume data to obtain the normalized unit historical traffic volume data, so as to improve the prediction accuracy of the taxi travel demand prediction model.

Specifically, since each unit historical traffic volume data has different dimensions and magnitudes, in order to avoid large differences in volume data between different units of historical traffic volume data, thereby affecting the reliability of the analysis results, normalization processing is performed on the historical traffic volume data, and the unit historical traffic volume data is scaled in proportion to fall into a small specific interval, thereby ensuring the reliability of the analysis results; the normalization processing of this embodiment adopts a deviation standardization method, that is, the minimum value in the data set is subtracted from each element value, and then divided by the difference between the maximum and minimum values in the data set, so that it falls within the interval [0,1], as shown in the following formula:

Wherein, _yk is the normalized unit historical traffic volume data of the kth unit interval, _tk is the unit historical traffic volume data of the kth unit interval (the element value in the data set to be converted), max(t) is the maximum value of the unit historical traffic volume data (i.e. the maximum value in the data set to be converted), min(t) is the minimum value of the unit historical traffic volume data (i.e. the minimum value in the data set to be converted); k is the number of the unit interval;

S5: Divide the area to be predicted into a spatial horizontal and vertical analysis matrix of n*n tuples, so that the normalized unit historical traffic volume data corresponds to each tuple, as an input set of the taxi travel demand prediction model. The rows of the spatial horizontal and vertical analysis matrix represent the longitude of the location of the area to be predicted, the columns of the spatial horizontal and vertical analysis matrix represent the latitude of the location of the area to be predicted, and the elements in the spatial horizontal and vertical analysis matrix are the unit historical traffic data of the corresponding unit interval;

Specifically, the spatial vertical and horizontal analysis matrix constructed in this embodiment divides each unit interval into a 250*250 matrix tuple form, and places the above-mentioned normalized historical traffic volume data in the corresponding position of the tuple as the input set of the taxi travel demand prediction model. Among them, the spatial vertical and horizontal analysis matrix represents the number of taxis at the location represented by the micro interval in the time period of the unit interval, but there are a total of 4032 matrices, and the rows and columns in the matrix represent the longitude and latitude respectively. The data in the matrix can represent the number of taxis at the longitude and latitude, that is, according to a certain element in the spatial vertical and horizontal analysis matrix, the number of taxis at what time (time) and where (longitude and latitude) can be known;

Specifically, in this embodiment, the historical traffic data of 4032 groups of taxis in Dalian within the historical time range threshold are collected by GPS installed on taxis. The normalized historical traffic data obtained after preprocessing and normalization are compressed into a 100*100 data matrix by np.zeros function, and the sliding window method is adopted. The loop function advances one unit each time to realize the sliding window operation, so that there is only one data sliding in the two adjacent data sets in the training data set. In addition, there is also one data sliding between the test data set and the training data set, which is conducive to mining the correlation in the trajectory sequence, thereby improving the accuracy. According to the time sequence, each 12 groups of data are classified, and 80% of the total data set are randomly selected as the training set (train_data) data set, and the other 20% are used as the test set (test_data) data set and output in npy format; and by appropriately compressing the dimension of the data set, it is compressed into a 100*100 data matrix, which improves the model prediction speed without affecting the accuracy.

In this embodiment, the input gate is a 100*100 matrix tuple after the normalization of the horizontal and vertical analysis matrix compressed by the np.zeros function. The matrix tuple contains the latitude and longitude, time, and normalized historical traffic volume data information; the forget gate determines how much of the unit state ct _-1 predicted at the previous moment is retained to the current moment _ct , which is model adaptation; the output gate controls the unit state to output the current prediction value, which is the predicted traffic volume data output at the last moment in this example.

S6. Obtaining a test set and a training set according to the normalized unit historical traffic data;

S7: According to the training set, the spatial horizontal and vertical analysis matrix is input into the taxi travel demand prediction model based on ConvLSTM to obtain the trained taxi travel demand prediction model based on ConvLSTM, the normalized historical traffic volume data is trained to obtain the predicted traffic volume data within the future time threshold range and the trained taxi travel demand prediction model based on ConvLSTM;

Preferably, the method for training the normalized historical traffic volume data according to the training set is as follows:

S71: Let i=1, x=0; wherein i is the row number of the spatial vertical and horizontal analysis matrix;

经过基于ConvLSTM的出租车出行需求预测模型 的卷积神经网络的输入层，其中S_i为空间纵横分析矩阵S中的第i行；

为空间纵横分析矩阵S中的第i行中的第n个数据；输入至所述卷积神经网 络的循环层，将所述循环层输出结果H_t-1与x_t分别卷积后，获得X＝[y,x]^T， 其中y为H_t-1经卷积后的数据，x为x_t经卷积后的数据；将其作为基于ConvLSTM 的出租车出行需求预测模型的输入，经过所述基于ConvLSTM的出租车出行需 求预测模型的卷积神经网络的LSTM后，获得预测交通量数据；其中，基于ConvLSTM的出租车出行需求预测模型的计算过程如公式(1)；S72:

After the input layer of the convolutional neural network based on the ConvLSTM taxi travel demand prediction model, where S _i is the i-th row in the spatial vertical and horizontal analysis matrix S;

is the nth data in the i-th row of the spatial vertical and horizontal analysis matrix S; input to the recurrent layer of the convolutional neural network, convolve the recurrent layer output result H _t-1 with x _t respectively, and obtain X = [y, x] ^T , wherein y is the data after convolution of H _t-1 , and x is the data after convolution of x _t ; use it as the input of the taxi travel demand prediction model based on ConvLSTM, and obtain the predicted traffic volume data after passing through the LSTM of the convolutional neural network of the taxi travel demand prediction model based on ConvLSTM; wherein, the calculation process of the taxi travel demand prediction model based on ConvLSTM is as shown in formula (1);

S8: input the test set into the trained ConvLSTM-based taxi travel demand prediction model to obtain test predicted traffic volume data; obtain the loss function between the test predicted traffic volume data and the real traffic volume data, and when the value of the loss function is less than the loss function threshold, execute S9, at which time the trained ConvLSTM-based taxi travel demand prediction model is the optimal ConvLSTM-based taxi travel demand prediction model; otherwise, repeat S7;

S9: Obtain the accuracy between the test predicted traffic volume data and the actual traffic volume data; and obtain the predicted traffic volume data within the future time threshold range according to the optimal ConvLSTM-based taxi travel demand prediction model.

Specifically, this embodiment uses the ConvLSTM-based taxi travel demand prediction model and uses the training data set to train the normalized historical traffic volume data to obtain the training predicted traffic volume data within the future time threshold range, imports the test set into the ConvLSTM-based taxi travel demand prediction model, tests the accuracy of the ConvLSTM-based taxi travel demand prediction model, and predicts the taxi traffic volume data within the future time threshold range;

Specifically, as shown in Figure 3, according to the taxi travel demand prediction model based on ConvLSTM, the output gate of the last layer of the ConvLSTM neural network is used as the prediction result of the entire network. The present invention predicts the taxi traffic volume at the city level by learning the historical taxi traffic volume data of Dalian, which greatly improves the accuracy of the prediction result.

Preferably, the loss function between the test predicted traffic volume data and the actual traffic volume data in S8 is: establishing a loss function to obtain the difference between the predicted traffic volume data obtained by the taxi travel demand prediction model based on ConvLSTM and the actual traffic volume data, and evaluating the taxi travel demand prediction model based on ConvLSTM;

Calculate the loss function loss based on the predicted value and the true value.

为基于ConvLSTM的出租车 出行需求预测模型的第i个预测值，output size为基于ConvLSTM的出租车出 行需求预测模型的参数个数，主要用来衡量基于ConvLSTM的出租车出行需求 预测模型所作出的预测离真实值之间的偏离程度，用以评估基于ConvLSTM的 出租车出行需求预测模型的预测结果的好坏。Where Loss is the loss function value; _yi is the i-th element value of the unit historical traffic volume data (i.e., data set) of the area to be predicted after normalization;

is the i-th predicted value of the taxi travel demand forecasting model based on ConvLSTM, and output size is the number of parameters of the taxi travel demand forecasting model based on ConvLSTM. It is mainly used to measure the deviation degree between the prediction made by the taxi travel demand forecasting model based on ConvLSTM and the true value, so as to evaluate the prediction result of the taxi travel demand forecasting model based on ConvLSTM.

Specifically, the loss function in this embodiment uses binary_crossentropy (i.e., binary cross entropy), and the accuracy uses the acc function to measure the degree of deviation between the predicted traffic volume data and the actual traffic volume data made by the taxi travel demand prediction model based on ConvLSTM, and evaluate the quality of the model prediction, wherein the learning rate is ɑ＝0.005, and the weight parameter and the bias parameter are initialized using the normal distribution (μ＝0.485, σ＝0.224), and the data batch of a single training process is 12. If the loss rate is greater than the set error value, the model is retrained, otherwise the model meets the requirements, and then the test data set is input into the trained ConvLSTM neural network to predict the future taxi travel demand density in Dalian.

Preferably: the accuracy between the test predicted traffic volume data and the actual traffic volume data in S9 is calculated as follows: the accuracy Acc is calculated based on the predicted value and the actual value,

为基于ConvLSTM的出租车出行需求预测模型的第 i个预测值更新参数重新计算预测值，ABS(·)表示取绝对值运算。In the formula, _yi is the i-th element value of the unit historical traffic volume data (i.e., data set) of the area to be predicted after normalization;

The parameters of the i-th prediction value of the taxi travel demand prediction model based on ConvLSTM are updated to recalculate the prediction value. ABS(·) represents the absolute value operation.

S9: According to the optimal ConvLSTM-based taxi travel demand prediction model, the predicted traffic volume data within the future time threshold range is obtained.

Specifically, after multiple rounds of training, the accuracy of this embodiment is shown in Figure 4. At this time, the accuracy has satisfied Acc<=0.01, so the model training is completed and meets the prediction requirements. The model at this time is the optimal ConvLSTM-based taxi travel demand prediction model. Finally, the taxi traffic volume in the entire Dalian city is predicted by the optimal ConvLSTM-based taxi travel demand prediction model, and the predicted traffic volume data of Dalian’s past taxi traffic is input as a certain time period in the future.

The present invention fully exploits the spatiotemporal correlation in the taxi trajectory sequence, and predicts the future traffic volume data of the taxi by learning the historical traffic volume data of the taxi, thereby improving the accuracy of traffic volume prediction. Traffic volume prediction refers to using the historical traffic volume data of the taxi, mainly GPS positioning data, to establish a prediction model based on a large amount of historical traffic volume data, using the known traffic volume distribution as the input of the prediction model, and obtaining the road where the taxi will be at the next moment through the deduction calculation of the model. Taxi traffic volume prediction can improve urban traffic safety and is part of intelligent transportation. For the traffic management department, it can effectively improve the control and utilization of taxis as a public transportation resource, and at the same time help to alleviate the congestion in the city, improve the utilization of urban roads and bring tangible benefits to the general public. However, the historical traffic volume data of taxi GPS has the characteristics of large data volume, high dimension, chaotic data, wide coverage, etc. It contains information such as taxi license plate number, passenger status, time, longitude, latitude, and speed in kilometers per hour. In addition, taxis are unevenly distributed in the spatiotemporal traffic network, and the correlation between taxi traffic volumes is relatively hidden and greatly affected by external factors, which makes it extremely difficult to predict taxi travel demand. If traditional prediction methods are used, this will inevitably cause great difficulties and the accuracy is also very low. Therefore, the prediction method based on deep learning is the only feasible method.

Based on this, the present invention discloses a ConvLSTM taxi travel demand prediction method, which solves the problem of mining the correlation of taxi trajectory data under complex external influencing factors, and improves the accuracy of the prediction model under massive taxi trajectory data. The established model is not only concise and highly operable, but also has a high accuracy rate.

The sliding window method is used in the training data set of the present invention, so that one data in two adjacent data sets moves, which not only increases the number of data sets, but also makes it easier to mine the correlation in the trajectory sequence, thereby improving the accuracy. And the ConvLSTM neural network used in the present invention can not only establish a temporal relationship like LSTM, but also can characterize local spatial features like CNN, greatly improving the accuracy of taxi trajectory prediction.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims (4)

1. A taxi travel demand prediction method based on ConvLSTM, characterized by comprising the following steps: S1: Establish a taxi travel demand prediction model based on ConvLSTM: S2. Obtain historical traffic volume data of the area to be predicted within the historical time range threshold, and pre-process the historical traffic volume data to obtain standard historical traffic volume data; the historical traffic volume data includes the taxi license plate number, passenger status, passenger time, longitude of the taxi location, latitude of the taxi location, and speed in kilometers per hour; S3, dividing the area to be predicted according to the directions of longitude and latitude to obtain a number of unit intervals, and obtaining unit historical traffic volume data of each unit interval within the historical time range threshold according to the standard historical traffic volume data; S4, normalizing the unit historical traffic volume data to obtain normalized unit historical traffic volume data; S5: Divide the area to be predicted into a spatial horizontal and vertical analysis matrix of n*n tuples, so that the normalized unit historical traffic volume data corresponds one-to-one to each tuple, the rows of the spatial horizontal and vertical analysis matrix represent the longitude of the location of the area to be predicted, the columns of the spatial horizontal and vertical analysis matrix represent the latitude of the location of the area to be predicted, and the elements in the spatial horizontal and vertical analysis matrix are the unit historical traffic data of the corresponding unit interval; S6. Obtaining a test set and a training set according to the normalized unit historical traffic volume data; S7: According to the training set, the spatial horizontal and vertical analysis matrix is input into the taxi travel demand prediction model based on ConvLSTM to obtain a trained taxi travel demand prediction model based on ConvLSTM; S8: Input the test set into the trained ConvLSTM-based taxi travel demand prediction model to obtain test predicted traffic volume data; obtain the loss function between the test predicted traffic volume data and the real traffic volume data, and when the value of the loss function is less than the loss function value threshold, execute S9, at which time the trained ConvLSTM-based taxi travel demand prediction model is the optimal ConvLSTM-based taxi travel demand prediction model; otherwise, repeat S7; S9: Obtain the accuracy rate between the test predicted traffic volume data and the actual traffic volume data; and obtain the predicted traffic volume data within the output future time threshold range according to the optimal ConvLSTM-based taxi travel demand prediction model. 2. According to the method for predicting taxi travel demand based on ConvLSTM in claim 1, it is characterized in that: in said S4, the normalized unit historical traffic volume data is obtained as follows:

Where _yk is the normalized unit historical traffic volume data of the kth unit interval, _tk is the unit historical traffic volume data of the kth unit interval, max(t) is the maximum value of the unit historical traffic volume data, min(t) is the minimum value of the unit historical traffic volume data; k is the number of the unit interval. 3. A taxi travel demand prediction method based on ConvLSTM according to claim 2, characterized in that: the loss function between the test predicted traffic volume data and the actual traffic volume data in S8 is:

Where Loss is the loss function value; _yi is the i-th element value of the unit historical traffic volume data of the area to be predicted after normalization;

is the i-th predicted value of the taxi travel demand prediction model based on ConvLSTM, and outputsize is the number of parameters of the taxi travel demand prediction model based on ConvLSTM. 4. A taxi travel demand prediction method based on ConvLSTM according to claim 3, characterized in that: the accuracy between the test predicted traffic volume data and the actual traffic volume data in S9 is calculated as follows:

Where _{yi is} the i-th element value of the unit historical traffic volume data of the area to be predicted after normalization;

is the i-th predicted value of the taxi travel demand prediction model based on ConvLSTM; ABS(·) represents the absolute value operation.

Images