Online car-hailing demand prediction method

Abstract

The invention discloses an online car-hailing demand prediction method, which comprises the following steps: firstly, constructing a prediction model based on a neural network, then collecting online car-hailing order data and air quality data of a place to be predicted in recent months, the air quality data comprising temperature, rainfall, PM2.5 index and air quality index (AQI), the method comprises the following steps: forming a plurality of feature matrixes according to collected online car-hailing order data and air quality data, forming a first training set by adopting the feature matrixes, and obtaining a daily periodic arrival flow quantity time sequence and a weekly periodic arrival flow quantity time sequence according to the collected online car-hailing order data, forming a second training set by adopting the daily periodic arrival flow quantity time sequence, forming a third training set by adopting the weekly periodic arrival flow quantity time sequence, training the prediction model by adopting the first training set, the second training set and the third training set to obtain a trained prediction model, and performing prediction through the trained prediction model; the method has the advantages of high prediction efficiency and high prediction accuracy.

Description

Technical Field

[0001] The present invention relates to online car-hailing demand forecasting technology, and in particular to an online car-hailing demand forecasting method. Background Art

[0002] In today's mobile internet information age, the development of smart transportation has greatly facilitated people's lives and significantly changed travel habits. Leveraging advanced science and technology, smart transportation enables intelligent management of urban traffic flow, including information collection and processing, as well as optimized dispatching, enabling efficient and effective transportation systems. As a key mode of transportation in urban public transportation, online ride-hailing services offer passengers precise and convenient travel services. However, the random nature of passenger travel in both time and space, coupled with drivers' blind search for passengers, results in uneven temporal and spatial distribution of online ride-hailing supply and demand. For example, areas with high demand during peak hours often experience a shortage of idle online ride-hailing vehicles and prolonged wait times. Drivers are unable to accurately locate high-demand areas, resulting in a large number of idle online ride-hailing vehicles remaining idle, hindering efficient passenger load factors. This mismatch between supply and demand can lead to regional capacity shortages, increased idle rates for online ride-hailing vehicles, and reduced road resource utilization.

[0003] In the field of forecasting research, numerous scholars have proposed various methods to improve the accuracy of travel demand forecasts. For example, one study established a deep learning framework based on ConvLSTM. This model uses geographic grid partitioning to input online ride-hailing demand data into a tabular form and extracts the spatiotemporal characteristics of online ride-hailing demand, thereby improving forecast accuracy. However, the ConvLSTM model has limitations when processing data over large spatial scales, resulting in low forecasting efficiency. Furthermore, existing online ride-hailing demand forecasting methods fail to fully consider the impact of factors such as temperature, rainfall, PM2.5 index, and air quality index on demand, ignoring the varying impacts of different factors on departure and arrival flows, leading to the risk of reduced forecast accuracy. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for predicting the demand for online ride-hailing vehicles with high prediction efficiency and high prediction accuracy.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a method for predicting the demand for online car-hailing, first building a prediction model based on a neural network, and then collecting the online car-hailing order data and air quality data of the predicted place in recent months, the air quality data including temperature, rainfall, PM2.5 index and air quality index (AQI), according to the collected online car-hailing order data and air quality data, obtaining the departure flow quantity time series data, arrival flow quantity time series data, temperature time series data, rainfall time series data, PM2.5 index time series data, air quality index time series data, daily periodic arrival flow quantity time series and weekly periodic arrival flow quantity time series, and using the departure flow quantity time series data, arrival flow quantity time series data, temperature time series data, rainfall time series data Sequence data, PM2.5 index time series data and air quality index time series data constitute multiple feature matrices, and these feature matrices are used to form a first training set, the daily periodic arrival flow number time series is used to form a second training set, and the weekly periodic arrival flow number time series is used to form a third training set; then the first training set, the second training set and the third training set are used to train the prediction model to obtain a trained prediction model. When predicting the demand for online car-hailing, the same construction method of the first training set, the second training set and the third training set is used to obtain the feature matrix of the place to be predicted in a preset previous period of time, the daily periodic arrival flow number time series and the weekly periodic arrival flow number time series, and they are input into the trained prediction model, and the trained prediction model generates a prediction result output for the demand for online car-hailing.

[0006] Compared with the existing technology, the advantage of the present invention is that it constructs a prediction model based on a neural network. When constructing a training set, the online car-hailing order data and air quality data of the predicted place in recent months are collected for training. Subsequently, it is not necessary to obtain data for a longer period of time for prediction, thereby reducing the computing resources required for data collection and processing and improving the prediction efficiency; and when training the prediction model, various factors including temperature, rainfall, PM2.5 index, air quality, and the internal characteristics of the departure flow and arrival flow are taken into account, so that the trained prediction model can establish the association between various factors including temperature, rainfall, PM2.5 index, air quality and online car-hailing demand, as well as the internal characteristics of the departure flow and arrival flow. In subsequent predictions, by constructing a feature matrix containing these factors for prediction, more accurate prediction results can be obtained, thereby improving the accuracy of the prediction.

[0007] Furthermore, the online car-hailing order data and air quality data of the predicted location in the past five months are collected.

[0008] Furthermore, based on the collected online ride-hailing order data and air quality data, the specific process of obtaining the departure flow quantity time series data, arrival flow quantity time series data, temperature time series data, rainfall time series data, PM2.5 index time series data, air quality index time series data, daily periodic arrival flow quantity time series data, and weekly periodic arrival flow quantity time series data is as follows:

[0009] Step 1: Divide the online ride-hailing order data from the past five months into multiple time periods, with each time period being 10 minutes. This yields a total of 26,496 time periods. Furthermore, the air quality data for the location to be predicted from the past five months is collected at 10-minute intervals, yielding a total of 26,496 temperature, rainfall, PM2.5, and air quality index (AQI) data points.

[0010] Step 2: Obtain the number of departure flows and arrival flows from the online ride-hailing order data for each time period. The departure flow refers to the number of online ride-hailing order data that departs from the location to be predicted, and the arrival flow refers to the number of online ride-hailing order data that ends at the location to be predicted.

[0011] Step 3: Arrange the number of departures in the online car-hailing order data for all time periods in chronological order to form departure flow time series data, and arrange the number of arrivals in the online car-hailing order data for all time periods in chronological order to form arrival flow time series data; arrange the collected temperature data in chronological order to form temperature time series data, arrange the collected rainfall data in chronological order to form rainfall time series data, arrange the collected PM2.5 index data in chronological order to form PM2.5 index time series data, and arrange the collected air quality index data in chronological order to form air quality index time series data;

[0012] Step 4: Take one day as a cycle, each cycle includes 144 time periods, and each group of 24 time periods in sequence constitutes a time series. The number of arrival flows in 26,496 time periods is periodically extracted to form a daily periodic arrival flow number time series; take one week as a cycle, each cycle includes 1008 time periods, and each group of 24 time periods in sequence constitutes a time series. The number of arrival flows in 26,496 time periods is periodically extracted to form a weekly periodic arrival flow number time series.

[0013] Furthermore, the number of departure flows in the jth time period is recorded as s j , the number of arrival flows in the jth time period is recorded as e j, record the pth air quality data in the jth time period as The first air quality data is rainfall data, the second air quality data is temperature data, the third air quality data is PM2.5 index data, and the fourth air quality data is air quality index data, j = 1, 2, ..., 26,496, p = 1, 2, 3, 4;

[0014] The first training set includes 25488 feature matrices, and the bth feature matrix is recorded as 3…25488, It is expressed by formula (1):

[0015]

[0016] Each feature matrix is used as a sample of the first training set. The data in rows 2 to 5 of each feature matrix are used as the predicted data of the sample, and the data in rows 1 and 6 are used as the label of the sample.

[0017] Furthermore, the second training set is recorded as E D , in, is the time series of the periodic arrival flow quantity on the bth day,

[0018] The third training set is denoted as E w , is the time series of the periodic arrival flow quantity in the bth week,

[0019] Furthermore, the prediction model includes a CNN neural network, an LSTM neural network, two BiLSTM neural networks, an attention module, a feature fusion layer and a fully connected layer. The CNN neural network is used to receive a feature matrix composed of departure flow quantity time series data, arrival flow quantity time series data, temperature time series data, rainfall time series data, PM2.5 index time series data and air quality index time series data, and capture the spatial characteristics of the data in the feature matrix to obtain spatial feature time series data that are output to the LSTM neural network and the attention module respectively. The LSTM neural network is used to capture the changes in the time dimension of the spatial feature time series data output by the CNN neural network to obtain the time dimension feature output to the attention module. The attention module is used to dynamically adjust the weight of the spatial feature time series data output by the LSTM neural network to obtain a weighted time dimension feature. The degree feature is output to the feature fusion layer, the first BiLSTM neural network is used to access the daily periodic arrival flow number time series, and capture the feature vector of the daily periodic arrival flow number time series, to obtain the daily periodic feature vector and output it to the feature fusion layer, the second BiLSTM neural network is used to access the weekly periodic arrival flow number time series, and capture the feature vector of the weekly periodic arrival flow number time series, to obtain the weekly periodic feature vector and output it to the feature fusion layer, the feature fusion layer is used to perform feature fusion on the weighted time dimension feature output by the attention module, the daily periodic feature vector output by the first BiLSTM neural network, and the weekly periodic feature vector output by the second BILSTM neural network, to obtain a comprehensive feature vector and output it to the fully connected layer, the fully connected layer is used to perform feature fusion and information integration on the comprehensive feature vector output by the feature fusion layer to generate a prediction result output.

[0020] Furthermore, when predicting the demand for online ride-hailing services at a certain location, the online ride-hailing service order data for the week before the current time of the location is divided into multiple time periods with each 10-minute period as a time period. A total of 1008 time periods of online ride-hailing service order data are obtained. The number of arrival flows in the online ride-hailing service order data for each time period is obtained, and the number of arrival flows in the tth time period is recorded as e' t , t = 1, 2, ..., 1088, where the 1008 time periods are arranged in sequence from 10 minutes before the current moment, and the 10 minutes after the current moment is the 1088th time period;

[0021] Taking one day as a cycle, each cycle includes 144 time periods. In order, each group of 24 time periods constitutes a time series. The number of arrival flows in 1008 time periods is periodically extracted to form a daily periodic arrival flow number time series. Taking one week as a cycle, each cycle includes 1008 time periods. In order, each group of 24 time periods constitutes a time series. The number of arrival flows in 1008 time periods is periodically extracted to form a weekly periodic arrival flow number time series.

[0022] The air quality data of the place for one week before the current moment is collected at 10-minute intervals, and an air quality data is collected every 10 minutes, so that a total of 1008 moments of temperature data, rainfall data, PM2.5 index data and air quality index (AQI) data are obtained; the pth air quality data at the tth moment is recorded as The data at the current moment is the data at the last moment, that is, the data at the 1088th moment;

[0023] The feature matrix of the place to be predicted is recorded as X' m , X' m It is expressed by formula (2):

[0024]

[0025] The time series of daily periodic arrival flow quantity is recorded as E d , E d It is expressed as follows using formula (3):

[0026] E d =[e' 864 ,e' 863 ,...,e' 843 ,e' 842 ,e' 841 ] (3)

[0027] The weekly periodic arrival flow number time series is recorded as E w , E w It is expressed as follows using formula (4):

[0028] E w =[e' 24 ,...,e'3,e'2,e'1] (4)

[0029] The feature matrix X' m Input CNN neural network, daily periodic arrival flow number time series E d Input the first Bilstm neural network, the weekly periodic arrival flow number time series E w Input the second Bilstm neural network to get the prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a structural diagram of the prediction model of the online car-hailing demand prediction method of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be described in further detail below with reference to the accompanying drawings and embodiments.

[0032] Example 1: A method for predicting demand for online car-hailing, first building a prediction model based on a neural network, and then collecting online car-hailing order data and air quality data of the predicted place in recent months, wherein the air quality data includes temperature, rainfall, PM2.5 index and air quality index (AQI), and obtaining departure flow quantity time series data, arrival flow quantity time series data, temperature time series data, rainfall time series data, PM2.5 index time series data, air quality index time series data, daily periodic arrival flow quantity time series and weekly periodic arrival flow quantity time series based on the collected online car-hailing order data and air quality data. .5 index time series data and air quality index time series data constitute multiple feature matrices, and these feature matrices are used to form the first training set, the daily periodic arrival flow number time series is used to form the second training set, and the weekly periodic arrival flow number time series is used to form the third training set; then the first training set, the second training set and the third training set are used to train the prediction model to obtain the trained prediction model. When predicting the demand for online car-hailing, the same construction method of the first training set, the second training set and the third training set is used to obtain the feature matrix of the place to be predicted in the preset previous period of time, the daily periodic arrival flow number time series and the weekly periodic arrival flow number time series, and they are input into the trained prediction model. The trained prediction model generates a prediction result output for the demand for online car-hailing.

[0033] Example 2: This example is basically the same as Example 1, except that in this example, the online car-hailing order data and air quality data of the predicted location in the last five months are collected.

[0034] In this embodiment, the specific process of obtaining departure flow quantity time series data, arrival flow quantity time series data, temperature time series data, rainfall time series data, PM2.5 index time series data, air quality index time series data, daily periodic arrival flow quantity time series data, and weekly periodic arrival flow quantity time series data based on the collected online ride-hailing order data and air quality data is as follows:

[0035] Step 1: Divide the online ride-hailing order data from the past five months into multiple time periods, with each time period being 10 minutes. This yields a total of 26,496 time periods. Furthermore, the air quality data for the location to be predicted from the past five months is collected at 10-minute intervals, yielding a total of 26,496 temperature, rainfall, PM2.5, and air quality index (AQI) data points.

[0036] Step 2: Obtain the number of departure flows and arrival flows from the online ride-hailing order data for each time period. The departure flow refers to the number of online ride-hailing order data that departs from the location to be predicted, and the arrival flow refers to the number of online ride-hailing order data that ends at the location to be predicted.

[0037] Step 3: Arrange the number of departures in the online car-hailing order data for all time periods in chronological order to form departure flow time series data, and arrange the number of arrivals in the online car-hailing order data for all time periods in chronological order to form arrival flow time series data; arrange the collected temperature data in chronological order to form temperature time series data, arrange the collected rainfall data in chronological order to form rainfall time series data, arrange the collected PM2.5 index data in chronological order to form PM2.5 index time series data, and arrange the collected air quality index data in chronological order to form air quality index time series data;

[0038] Step 4: Take one day as a cycle, each cycle includes 144 time periods, and each group of 24 time periods in sequence constitutes a time series. The number of arrival flows in 26,496 time periods is periodically extracted to form a daily periodic arrival flow number time series; take one week as a cycle, each cycle includes 1008 time periods, and each group of 24 time periods in sequence constitutes a time series. The number of arrival flows in 26,496 time periods is periodically extracted to form a weekly periodic arrival flow number time series.

[0039] Example 3: This example is basically the same as Example 2, except that in this example, the number of departure flows in the jth time period is recorded as s j , the number of arrival flows in the jth time period is recorded as e j , record the pth air quality data in the jth time period as The first air quality data is rainfall data, the second air quality data is temperature data, the third air quality data is PM2.5 index data, and the fourth air quality data is air quality index data, j = 1, 2, ..., 26,496, p = 1, 2, 3, 4;

[0040] The first training set includes 25488 feature matrices, and the bth feature matrix is recorded as It is expressed by formula (1):

[0041]

[0042] Each feature matrix is used as a sample of the first training set. The data in rows 2 to 5 of each feature matrix are used as the predicted data of the sample, and the data in rows 1 and 6 are used as the label of the sample.

[0043] 5. A method for predicting online car-hailing demand according to claim 4, characterized in that the second training set is denoted as E D , in, is the time series of the periodic arrival flow quantity on the bth day,

[0044] The third training set is denoted as E w , is the time series of the periodic arrival flow quantity in the bth week,

[0045] In this example, ride-hailing order data is divided into 10-minute time periods, and air quality data is acquired in 10-minute time periods, resulting in a total of 26,496 time periods. This data acquisition method can more accurately capture the dynamic changes in ride-hailing demand, especially demand fluctuations during peak hours and under special weather conditions, thereby improving forecast accuracy. The number of departures and arrivals, as well as rainfall, temperature, PM2.5 index, and air quality index, are integrated into the feature matrix, allowing the training set to incorporate multiple factors and generate more accurate forecast results. This example also incorporates daily and weekly periodic arrival time series. This means that the dataset construction takes into account the cyclical patterns of ride-hailing demand within a day and a week, thereby improving forecast accuracy. For example, during the morning and evening rush hours on weekdays, ride-hailing demand typically increases significantly, while weekend demand patterns may differ from weekdays. This allows the prediction model to be trained with predictive capabilities across a variety of scenarios, improving its accuracy.

[0046] Example 4: This example is basically the same as Example 3, except that: in this example, Figure 1As shown, the prediction model includes a CNN neural network, an LSTM neural network, two BiLSTM neural networks, an attention module, a feature fusion layer and a fully connected layer. The CNN neural network is used to receive a feature matrix composed of departure flow quantity time series data, arrival flow quantity time series data, temperature time series data, rainfall time series data, PM2.5 index time series data and air quality index time series data, and capture the spatial features of the data in the feature matrix to obtain spatial feature time series data, which are output to the LSTM neural network and the attention module respectively. The LSTM neural network is used to capture the changes in the time dimension of the spatial feature time series data output by the CNN neural network to obtain the time dimension feature output to the attention module. The attention module is used to dynamically adjust the weight of the spatial feature time series data output by the LSTM neural network to obtain the weighted time dimension feature. Output to the feature fusion layer, the first BiLSTM neural network is used to access the daily periodic arrival flow number time series, and capture the feature vector of the daily periodic arrival flow number time series, obtain the daily periodic feature vector and output it to the feature fusion layer, the second BiLSTM neural network is used to access the weekly periodic arrival flow number time series, and capture the feature vector of the weekly periodic arrival flow number time series, obtain the weekly periodic feature vector and output it to the feature fusion layer, the feature fusion layer is used to perform feature fusion on the weighted time dimension features output by the attention module, the daily periodic feature vector output by the first BiLSTM neural network, and the weekly periodic feature vector output by the second BILSTM neural network, obtain a comprehensive feature vector and output it to the fully connected layer, the fully connected layer is used to perform feature fusion and information integration on the comprehensive feature vector output by the feature fusion layer to generate a prediction result output.

[0047] In this embodiment, the prediction model is a hybrid model that combines a CNN neural network, an LSTM neural network, two BiLSTM neural networks, an attention module, a feature fusion layer, and a fully connected layer. The hybrid model performs predictions in three ways and then combines the three prediction results to obtain the final prediction result. It can more comprehensively capture and integrate various features that affect the demand for online ride-hailing services, and generate more accurate prediction results.

[0048] Example 5: This example is basically the same as Example 4, except that: in this example, when predicting the demand for online car-hailing at a certain location, the online car-hailing order data for the week before the current time of the location is divided into multiple time periods of online car-hailing order data with every 10 minutes as a time period. A total of 1008 time periods of online car-hailing order data are obtained. The number of arrival flows in the online car-hailing order data of each time period is obtained respectively, and the number of arrival flows in the t-th time period is recorded as e' t , t = 1, 2, ..., 1008, where the 1008 time periods are arranged in sequence from 10 minutes before the current moment, and the 10 minutes after the current moment is the 1008th time period;

[0049] Taking one day as a cycle, each cycle includes 144 time periods. In order, each group of 24 time periods constitutes a time series. The number of arrival flows in 1008 time periods is periodically extracted to form a daily periodic arrival flow number time series. Taking one week as a cycle, each cycle includes 1008 time periods. In order, each group of 24 time periods constitutes a time series. The number of arrival flows in 1008 time periods is periodically extracted to form a weekly periodic arrival flow number time series.

[0050] The air quality data of the place for one week before the current moment is collected at 10-minute intervals, and an air quality data is collected every 10 minutes, so that a total of 1008 moments of temperature data, rainfall data, PM2.5 index data and air quality index (AQI) data are obtained; the pth air quality data at the tth moment is recorded as The data at the current moment is the data at the last moment, that is, the data at the 1008th moment;

[0051] The feature matrix of the place to be predicted is recorded as X' m , X' m It is expressed by formula (2):

[0052]

[0053] The time series of daily periodic arrival flow quantity is recorded as E d , E d It is expressed as follows using formula (3):

[0054] E d =[e' 864 ,e' 863 ,…,e' 843 ,e' 842 ,e' 841 ] (3)

[0055] The weekly periodic arrival flow number time series is recorded as Ew , E w It is expressed as follows using formula (4):

[0056] E w =[e' 24 ,...,e'3,e'2,e'1] (4)

[0057] The feature matrix X' m Input CNN neural network, daily periodic arrival flow number time series E d Input the first Bilstm neural network, the weekly periodic arrival flow number time series E w Input the second Bilstm neural network to get the prediction result.

[0058] In this embodiment, only the most recent week of data is needed to make a prediction, eliminating the need for data spanning longer time periods. This significantly reduces the computing resources required for data collection and processing, effectively improving prediction efficiency. The hybrid neural network architecture, combining a CNN neural network, an LSTM neural network, two BiLSTM neural networks, an attention module, a feature fusion layer, and a fully connected layer, is employed to more comprehensively capture and integrate the various features that influence ride-hailing demand, generating more accurate prediction results. Furthermore, the present invention considers multiple factors, including temperature, rainfall, PM2.5 index, and air quality index, as well as the internal characteristics of departure and arrival flows. By constructing a feature matrix encompassing these factors, the accuracy of the prediction is further improved.

[0059] To verify the performance of our proposed method for predicting online ride-hailing demand, we set the city center of Haikou as the location for prediction. We collected online ride-hailing order data from May 10, 2017, to May 16, 2017, and analyzed the departure and arrival traffic. We also collected air quality data for the period, measured in 10-minute time intervals, as shown in Table 1.

[0060] Table 1

[0061]

[0062]

[0063] According to the method of Example 5, the final predicted demand for departure traffic in the center of Haikou during the current forecast period is 1874, and the demand for arrival traffic is 1673. By querying the records, the actual departure traffic used in the center of Haikou during this period is 1873, and the arrival traffic is 1676, which shows that the prediction results of the online ride-hailing demand forecasting method of the present invention are highly accurate.

[0064] In summary, the online car-hailing demand forecasting method of the present invention adopts a hybrid neural network architecture, combining a CNN neural network, an LSTM neural network, two BiLSTM neural networks, an attention module, a feature fusion layer and a fully connected layer to build a forecasting model. During training, the training samples include a variety of factors including temperature, rainfall, PM2.5 index, air quality index, and the internal characteristics of the departure flow and arrival flow, so that the forecasting model can more comprehensively capture and integrate the various characteristics that affect the demand for online car-hailing, and generate more accurate forecasting results. In addition, the present invention only needs to collect data from the past week to make forecasts, without the need to obtain data for a longer time span, thereby significantly reducing the computing resources required for data collection and processing, and effectively improving the forecasting efficiency.

Claims

1. A method for predicting demand for online ride-hailing services, characterized in that First, a prediction model is constructed based on a neural network. Then, the online car-hailing order data and air quality data of the predicted place in recent months are collected. The air quality data include temperature, rainfall, PM2.5 index and air quality index (AQI). According to the collected online car-hailing order data and air quality data, the departure flow quantity time series data, the arrival flow quantity time series data, the temperature time series data, the rainfall time series data, the PM2.5 index time series data, the air quality index time series data, the daily periodic arrival flow quantity time series and the weekly periodic arrival flow quantity time series are obtained. The departure flow quantity time series data, the arrival flow quantity time series data, the temperature time series data, the rainfall time series data, the PM2.5 index time series data are used. The data and air quality index time series data constitute multiple feature matrices, and these feature matrices are used to form a first training set, the daily periodic arrival flow number time series is used to form a second training set, and the weekly periodic arrival flow number time series is used to form a third training set; then the first training set, the second training set and the third training set are used to train the prediction model to obtain a trained prediction model. When predicting the demand for online car-hailing, the same construction method of the first training set, the second training set and the third training set is used to obtain the feature matrix of the place to be predicted in a preset previous period of time, the daily periodic arrival flow number time series and the weekly periodic arrival flow number time series, and they are input into the trained prediction model, and the trained prediction model generates a prediction result output for the demand for online car-hailing.

2. A method for predicting online car-hailing demand according to claim 1, characterized in that The data collected are the online taxi order data and air quality data of the predicted location in the past five months.

3. A method for predicting online car-hailing demand according to claim 2, characterized in that The specific process of obtaining the departure flow quantity time series data, arrival flow quantity time series data, temperature time series data, rainfall time series data, PM2.5 index time series data, air quality index time series data, daily periodic arrival flow quantity time series data, and weekly periodic arrival flow quantity time series data based on the collected online ride-hailing order data and air quality data is as follows: Step 1: Divide the online ride-hailing order data from the past five months into multiple time periods, with each time period being 10 minutes. This yields a total of 26,496 time periods. Furthermore, the air quality data for the location to be predicted from the past five months is collected at 10-minute intervals, yielding a total of 26,496 temperature, rainfall, PM2.5, and air quality index (AQI) data points. Step 2: Obtain the number of departure flows and arrival flows from the online ride-hailing order data for each time period. The departure flow refers to the number of online ride-hailing order data that departs from the location to be predicted, and the arrival flow refers to the number of online ride-hailing order data that ends at the location to be predicted. Step 3: Arrange the number of departures in the online car-hailing order data for all time periods in chronological order to form departure flow time series data, and arrange the number of arrivals in the online car-hailing order data for all time periods in chronological order to form arrival flow time series data; arrange the collected temperature data in chronological order to form temperature time series data, arrange the collected rainfall data in chronological order to form rainfall time series data, arrange the collected PM2.5 index data in chronological order to form PM2.5 index time series data, and arrange the collected air quality index data in chronological order to form air quality index time series data; Step 4: Take one day as a cycle, each cycle includes 144 time periods, and each group of 24 time periods in sequence constitutes a time series. The number of arrival flows in 26,496 time periods is periodically extracted to form a daily periodic arrival flow number time series; take one week as a cycle, each cycle includes 1008 time periods, and each group of 24 time periods in sequence constitutes a time series. The number of arrival flows in 26,496 time periods is periodically extracted to form a weekly periodic arrival flow number time series.

4. A method for predicting online car-hailing demand according to claim 3, characterized in that The number of departure flows in the jth time period is recorded as s j , the number of arrival flows in the jth time period is recorded as e j , record the pth air quality data in the jth time period as The first air quality data is rainfall data, the second air quality data is temperature data, the third air quality data is PM2.5 index data, and the fourth air quality data is air quality index data, j = 1, 2, ..., 26,496, p = 1, 2, 3, 4; The first training set includes 25488 feature matrices, and the bth feature matrix is recorded as b=1,2,3…25488, It is expressed by formula (1): Each feature matrix is used as a sample of the first training set. The data in rows 2 to 5 of each feature matrix are used as the predicted data of the sample, and the data in rows 1 and 6 are used as the label of the sample.

5. A method for predicting online car-hailing demand according to claim 4, characterized in that The second training set is denoted as E D , in, is the time series of the periodic arrival flow quantity on the bth day, The third training set is denoted as E w , is the time series of the periodic arrival flow quantity in the bth week, 6. A method for predicting online car-hailing demand according to claim 3, characterized in that The prediction model includes a CNN neural network, an LSTM neural network, two BiLSTM neural networks, an attention module, a feature fusion layer and a fully connected layer. The CNN neural network is used to receive a feature matrix composed of departure flow quantity time series data, arrival flow quantity time series data, temperature time series data, rainfall time series data, PM2.5 index time series data and air quality index time series data, and capture the spatial features of the data in the feature matrix to obtain spatial feature time series data, which are output to the LSTM neural network and the attention module respectively. The LSTM neural network is used to capture the changes in the time dimension of the spatial feature time series data output by the CNN neural network to obtain the time dimension features output to the attention module. The attention module is used to dynamically adjust the weight of the spatial feature time series data output by the LSTM neural network to obtain the weighted time dimension features. Output to the feature fusion layer, the first BiLSTM neural network is used to access the daily periodic arrival flow number time series, and capture the feature vector of the daily periodic arrival flow number time series, obtain the daily periodic feature vector and output it to the feature fusion layer, the second BiLSTM neural network is used to access the weekly periodic arrival flow number time series, and capture the feature vector of the weekly periodic arrival flow number time series, obtain the weekly periodic feature vector and output it to the feature fusion layer, the feature fusion layer is used to perform feature fusion on the weighted time dimension features output by the attention module, the daily periodic feature vector output by the first BiLSTM neural network, and the weekly periodic feature vector output by the second BILSTM neural network, obtain a comprehensive feature vector and output it to the fully connected layer, the fully connected layer is used to perform feature fusion and information integration on the comprehensive feature vector output by the feature fusion layer to generate a prediction result output.

7. A method for predicting online car-hailing demand according to claim 6, characterized in that When predicting the demand for online ride-hailing services at a certain location, the online ride-hailing service order data for the week before the current time of the location is divided into multiple time periods with each 10-minute period as a time period. A total of 1008 time periods of online ride-hailing service order data are obtained. The number of arrival flows in the online ride-hailing service order data for each time period is obtained, and the number of arrival flows in the tth time period is recorded as e' t , t = 1, 2, ..., 1088, where the 1008 time periods are arranged in sequence from 10 minutes before the current moment, and the 10 minutes after the current moment is the 1088th time period; Taking one day as a cycle, each cycle includes 144 time periods. In order, each group of 24 time periods constitutes a time series. The number of arrival flows in 1008 time periods is periodically extracted to form a daily periodic arrival flow number time series. Taking one week as a cycle, each cycle includes 1008 time periods. In order, each group of 24 time periods constitutes a time series. The number of arrival flows in 1008 time periods is periodically extracted to form a weekly periodic arrival flow number time series. The air quality data of the place for one week before the current moment is collected at 10-minute intervals, and an air quality data is collected every 10 minutes, so that a total of 1008 moments of temperature data, rainfall data, PM2.5 index data and air quality index (AQI) data are obtained; the pth air quality data at the tth moment is recorded as t=1, 2, ..., 1088, the data at the current moment is the data at the last moment, that is, the data at the 1088th moment; The feature matrix of the place to be predicted is recorded as X' m , X' m It is expressed by formula (2): The time series of daily periodic arrival flow quantity is recorded as E d , E d It is expressed as follows using formula (3): AND d =[and' 864 ,And' 863 ,...,And' 843 ,And' 842 ,And' 841 ] (3) The weekly periodic arrival flow number time series is recorded as E w , E w It is expressed as follows using formula (4): AND w =[and' 24 ,...,e'3,e'2,e'1] (4) The feature matrix X' m Input CNN neural network, daily periodic arrival flow number time series E d Input the first Bilstm neural network, the weekly periodic arrival flow number time series E w Input the second Bilstm neural network to get the prediction result.

Images