Joint forecasting method of taxi and online car-hailing demand

Abstract

The present invention specifically relates to a joint demand prediction method based on taxis and online car-hailing vehicles, including: inputting taxi historical demand matrices and online car-hailing vehicle historical demand matrices into a demand prediction model, and outputting corresponding future demand prediction values; firstly generating taxi spatiotemporal sharing information and online car-hailing vehicle spatiotemporal sharing information; then extracting spatiotemporal features based on the taxi historical demand matrix and the adaptive adjacency matrix, and generating a taxi future demand matrix in combination with the online car-hailing vehicle spatiotemporal sharing information; at the same time, extracting spatiotemporal features based on the online car-hailing vehicle historical demand matrix and the adaptive adjacency matrix, and generating a online car-hailing vehicle future demand matrix in combination with the taxi spatiotemporal sharing information; finally generating a rental car future demand prediction value and an online car-hailing vehicle future demand prediction value. The present invention can extract intra-mode features and inter-mode features of taxis and online car-hailing vehicles, and can realize the fusion of intra-mode features and inter-mode features, thereby realizing joint demand prediction of taxis and online car-hailing vehicles.

Description

Technical Field

[0001] The present invention relates to the field of traffic demand forecasting, and in particular to a joint forecasting method for demand based on taxis and online car-hailing services. Background Art

[0002] The increasing demand for travel caused by overcrowding has brought tremendous pressure and severe challenges to urban transportation systems. In response to the increasing difficulty of travel, a variety of innovative travel modes have been launched and dominated. The most representative is the mobility on demand (MoD) service provided by online ride-hailing companies such as Uber, Lyft, and Didi, which provides passengers with flexible and convenient travel options. However, the problem of travel difficulties has not been solved, or even worse, mainly due to the imbalance between transportation supply and demand. Whether it is people (i.e. taxi drivers) or machines (i.e. online ride-hailing platforms), timely access to accurate information on both transportation supply and passenger demand is the basis and necessary condition for making the best online decisions.

[0003] In actual situations, it is quite challenging to understand demand compared to supply information. Supply information can be easily obtained and monitored through the GPS positioning and vacancy status records of vehicles, while passenger demand cannot be naturally sensed but can only be predicted. Therefore, an accurate and timely passenger demand forecasting model is a key driver. On the one hand, the forecast results can help taxi drivers avoid staying in low-demand areas and go to high-demand areas. On the other hand, online ride-hailing platforms can also benefit from forecasting demand. For example, based on the forecasted demand, the platform can better allocate and dispatch vehicle resources in advance to achieve a win-win situation, which can not only reduce the driver's efforts to find passengers, but also reduce the waiting time of passengers.

[0004] Taxi services and online ride-hailing services are completely different but related modes of transportation (modes). However, most existing studies focus on predicting the demand of a single mode. The premise of these prediction models is that a single mode of transportation service is a fairly independent system, and future demand is only related to data-driven historical observed demand. In fact, passengers are usually not bound to one mode of transportation. They can switch between different modes as needed. Although there is mode switching behavior, it is more intensive and frequent between taxis and online ride-hailing services. For example, when a passenger waits too long for a taxi, he / she may turn to online ride-hailing services, but rarely look for buses around. Conversely, due to the platform's surge pricing strategy, if the dynamic price is too high, online ride-hailing passengers may choose to take a taxi to save money. These mode switching behaviors show that passengers have different demands for taxis and online ride-hailing services, but they influence each other. In other words, the above premise is not always reasonable. Moreover, these influences exist dynamically in both spatial and temporal dimensions, and form complex relationships between the demands of various modes. The applicant believes that the demand forecast for one mode should not only consider its historical demand, but also the dynamic changes in the demands of other modes. That is, demand forecasting for taxi mode and ride-hailing mode should be combined and regarded as two related tasks.

[0005] Therefore, how to design a method that can achieve joint prediction of taxi and online car-hailing demand is a technical problem that needs to be solved urgently. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the technical problem to be solved by the present invention is: how to provide a joint demand prediction method based on taxis and online-hailing cars, which can extract the intra-mode features and inter-mode features of taxis and online-hailing cars, and can realize the fusion of intra-mode features and inter-mode features, thereby realizing the joint demand prediction of taxis and online-hailing cars, thereby improving the accuracy and comprehensiveness of traffic demand prediction, and providing a new idea for the joint demand prediction of multiple modes of transportation.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] The joint demand forecasting method based on taxis and online ride-hailing services includes:

[0009] S1: Obtain the corresponding taxi historical demand matrix and online car-hailing historical demand matrix;

[0010] S2: Input the taxi historical demand matrix and the online car-hailing historical demand matrix into the trained demand forecasting model, and output the corresponding taxi future demand forecast value and online car-hailing future demand forecast value;

[0011] The demand forecasting model first generates the corresponding taxi spatiotemporal sharing information and online car-hailing spatiotemporal sharing information based on the taxi historical demand matrix and the online car-hailing historical demand matrix combined with the corresponding adaptive adjacency matrix; then, the spatiotemporal features are extracted based on the taxi historical demand matrix and the corresponding adaptive adjacency matrix, and the corresponding taxi future demand matrix is ​​generated in combination with the online car-hailing spatiotemporal sharing information; at the same time, the spatiotemporal features are extracted based on the online car-hailing historical demand matrix and the corresponding adaptive adjacency matrix, and the corresponding online car-hailing future demand matrix is ​​generated in combination with the taxi spatiotemporal sharing information; finally, the taxi future demand forecast value and the online car-hailing future demand forecast value are generated based on the taxi future demand matrix and the online car-hailing future demand matrix;

[0012] S3: The future demand forecast value of taxis and the future demand forecast value of online ride-hailing vehicles are used as the joint demand forecast result of taxis and online ride-hailing vehicles.

[0013] Preferably, the demand forecasting model includes:

[0014] The input layer is used to obtain the historical demand matrix of taxis and the historical demand matrix of online ride-hailing vehicles;

[0015] A spatiotemporal interweaving component is used to generate corresponding taxi spatiotemporal sharing information and online car-hailing spatiotemporal sharing information according to the taxi historical demand matrix and the online car-hailing historical demand matrix in combination with the corresponding adaptive adjacency matrix;

[0016] The taxi component is used to extract spatiotemporal features based on the taxi historical demand matrix and the corresponding adaptive adjacency matrix, and integrate the spatiotemporal sharing information of online ride-hailing to generate the corresponding taxi future demand matrix;

[0017] The online car-hailing component is used to extract spatiotemporal features based on the online car-hailing historical demand matrix and the corresponding adaptive adjacency matrix, and integrate the taxi spatiotemporal sharing information to generate the corresponding online car-hailing future demand matrix;

[0018] The output layer is used to perform demand forecasting based on the taxi future demand matrix and the online car-hailing future demand matrix through a fully connected network, and generate and output the corresponding taxi future demand forecast value and online car-hailing future demand forecast value.

[0019] Preferably, the space-time interleaving component includes a time interleaving module and a space interleaving module;

[0020] Both the taxi component and the online car-hailing component include several spatiotemporal blocks; each spatiotemporal block includes corresponding temporal convolutional layers and spatial convolutional layers;

[0021] The time interleaving module is used to perform time convolution on the taxi historical demand matrix and the online car-hailing historical demand matrix respectively to generate corresponding taxi time sharing information and online car-hailing time sharing information;

[0022] For a single spatiotemporal block of a taxi component: extract temporal features based on the taxi historical demand matrix through a temporal convolutional layer, and integrate the online car-hailing time sharing information to generate the corresponding taxi demand temporal feature matrix;

[0023] For a single spatiotemporal block of the online car-hailing component: extract temporal features based on the online car-hailing historical demand matrix through the temporal convolutional layer, and integrate the taxi time sharing information to generate the corresponding online car-hailing demand time feature matrix;

[0024] The spatial interleaving module is used to extract spatial features based on the taxi demand time feature matrix and the online car-hailing demand time feature matrix combined with the corresponding adaptive adjacency matrix, and generate corresponding taxi spatiotemporal sharing information and online car-hailing spatiotemporal sharing information;

[0025] For a single spatiotemporal block of a taxi component: spatial features are extracted based on the taxi demand temporal feature matrix and the corresponding adaptive adjacency matrix through a spatial convolutional layer, and the spatiotemporal sharing information of online ride-hailing services is integrated to generate the corresponding taxi demand spatiotemporal feature matrix;

[0026] For a single spatiotemporal block of the online car-hailing component: spatial features are extracted based on the online car-hailing demand temporal feature matrix and the corresponding adaptive adjacency matrix through the spatial convolution layer, and the taxi spatiotemporal sharing information is integrated to generate the corresponding online car-hailing demand spatiotemporal feature matrix.

[0027] Preferably, the time interleaving module includes a first time convolution layer and a second time convolution layer for performing time convolution on the taxi historical demand matrix and the online car-hailing historical demand matrix respectively;

[0028] The temporal convolution formulas for the first temporal convolution layer and the second temporal convolution layer are as follows:

[0029]

[0030]

[0031] Where: They represent taxi time sharing information and online car-hailing time sharing information respectively; X TA , X RS They represent the historical demand matrix of taxis and the historical demand matrix of online ride-hailing cars respectively; TCL1 and TCL2 represent the first time convolution layer and the second time convolution layer respectively;

[0032] For a single spatiotemporal block of the taxi component: the taxi historical demand matrix is ​​used as the input of the temporal convolution layer and temporally convolved to generate the corresponding taxi demand temporal feature matrix; then the taxi demand temporal feature matrix and the online car-hailing time sharing information are fused in the last channel of the time dimension to obtain the final taxi demand temporal feature matrix;

[0033] The formula is described as:

[0034]

[0035]

[0036] Where: X′ TA represents the taxi demand time characteristic matrix; X TA represents the taxi historical demand matrix; F t represents the number of channels; L′ represents the length of the time dimension; X′ TA (:,L′,:), They represent the taxi demand time feature matrix and the last channel of the online car-hailing time sharing information in the time dimension respectively;

[0037] For a single spatiotemporal block of the online car-hailing component: the online car-hailing historical demand matrix is ​​used as the input of the temporal convolution layer and temporally convolved to generate the corresponding online car-hailing demand time feature matrix; then the online car-hailing demand time feature matrix and the taxi time sharing information are fused in the last channel of the time dimension to obtain the final online car-hailing demand time feature matrix;

[0038] The formula is described as:

[0039]

[0040]

[0041] Where: X′ RS represents the time characteristic matrix of online car-hailing demand; X RS represents the historical demand matrix of online ride-hailing; F t represents the number of channels; L′ represents the length of the time dimension; X′ RS (:,L′,:), They respectively represent the last channel of the online car-hailing demand time feature matrix and the taxi time sharing information in the time dimension.

[0042] Preferably, the time interleaving module filters the shared information through a self-gating mechanism combined with the following formula:

[0043]

[0044]

[0045] Where: They represent taxi time sharing information and online car-hailing time sharing information respectively; σ represents the Sigmod function; ⊙ represents the Hadamard product (element-wise product) operator.

[0046] Preferably, the spatial interleaving module includes a first spatial convolution layer and a second spatial convolution layer for performing spatial convolution on the taxi demand time feature matrix and the online car-hailing demand time feature matrix respectively;

[0047] The spatial convolution formulas of the first spatial convolution layer and the second spatial convolution layer are as follows:

[0048]

[0049]

[0050]

[0051]

[0052] Where: Respectively represent the time-space sharing information of taxis and online car-hailing; A RT , A TR Represent the adaptive adjacency matrix between taxi and online car-hailing demands, where A RT represents the relationship between the taxi demand in one area and the online car-hailing demand in other areas. TR represents the relationship between the online car-hailing demand in one area and the taxi demand in other areas; X′ TA , X′ RS They represent the taxi demand time characteristic matrix and the online car-hailing demand time characteristic matrix respectively; represents two feature transformation matrices; E Tq , represents the regional characteristics set for taxi demand; E Rq , represents the regional characteristics set for online ride-hailing demand; N represents the number of regions;

[0053] For a single spatiotemporal block of a taxi component: the taxi demand temporal feature matrix is ​​used as the input of the spatial convolution layer and combined with the corresponding adaptive adjacency matrix for spatial convolution to generate the corresponding taxi demand spatiotemporal feature matrix; then the taxi demand spatiotemporal feature matrix is ​​fused with the online car-hailing spatiotemporal sharing information to obtain the final taxi demand spatiotemporal feature matrix;

[0054] The formula is described as:

[0055]

[0056]

[0057] The graph learning sublayer is used to automatically discover the dependencies between different regions and generate the corresponding adaptive adjacency matrix.

[0058] The formula is described as:

[0059]

[0060] Where: X″ TA represents the spatiotemporal characteristic matrix of taxi demand; X′ TA A represents the taxi demand time characteristic matrix; TT An adaptive adjacency matrix representing taxi demand between N regions; represents the projection used for feature transformation; σ(·) represents the GLU activation function; F s Represents the output dimension of the graph convolutional layer; represents the time-space sharing information of online car-hailing; E Tq , represents the regional characteristics set for taxi demand; N represents the number of regions;

[0061] For a single spatiotemporal block of the online car-hailing component: the online car-hailing demand temporal feature matrix is ​​used as the input of the spatial convolution layer and combined with the corresponding adaptive adjacency matrix for spatial convolution to generate the corresponding online car-hailing demand spatiotemporal feature matrix; then the online car-hailing demand spatiotemporal feature matrix is ​​fused with the taxi spatiotemporal sharing information to obtain the final online car-hailing demand spatiotemporal feature matrix;

[0062] The formula is described as:

[0063]

[0064]

[0065] The graph learning sublayer is used to automatically discover the dependencies between different regions and generate the corresponding adaptive adjacency matrix.

[0066] The formula is described as:

[0067]

[0068] Where: X″ RS represents the spatiotemporal characteristic matrix of online car-hailing demand; X′ RS A represents the time characteristic matrix of online car-hailing demand; RR An adaptive adjacency matrix representing the demand for online ride-hailing services between N regions; represents the projection used for feature transformation; σ(·) represents the GLU activation function; Fs Represents the output dimension of the graph convolutional layer; represents the taxi space-time sharing information; E Rq , Represents the regional characteristics set for online ride-hailing demand; N represents the number of regions.

[0069] Preferably, a residual connection structure is provided in the spatiotemporal block;

[0070] For a single spatiotemporal block of the taxi component: the taxi historical demand matrix and the corresponding taxi demand spatiotemporal feature matrix are fused to generate the taxi demand matrix of the spatiotemporal block;

[0071] The formula is described as:

[0072]

[0073] Where: represents the taxi demand matrix of the b+1th spatiotemporal block in the taxi component; represents the taxi historical demand matrix of the b-th spatiotemporal block input in the taxi component; represents the taxi demand time characteristic matrix of the bth spatiotemporal block in the taxi component; represents the spatiotemporal characteristic matrix of taxi demand for the bth spatiotemporal block in the taxi component;

[0074] For a single spatiotemporal block of the online ride-hailing component: the online ride-hailing historical demand matrix and the corresponding online ride-hailing demand spatiotemporal feature matrix are merged to generate the online ride-hailing demand matrix of the spatiotemporal block;

[0075] The formula is described as:

[0076]

[0077] Where: represents the online car-hailing demand matrix of the b+1th spatiotemporal block in the online car-hailing component; Represents the historical demand matrix of online ride-hailing input in the b-th spatiotemporal block in the online ride-hailing component; represents the temporal convolution of the b-th spatiotemporal block in the online car-hailing component; Represents the spatial convolution of the b-th spatiotemporal block in the ride-hailing component.

[0078] Preferably, a jump connection structure is provided between the space-time blocks;

[0079] The taxi component and the online car-hailing component respectively connect the taxi demand matrices and online car-hailing demand matrices of all their time-space blocks together to generate the corresponding taxi future demand matrices and online car-hailing future demand matrices.

[0080] Preferably, the output layer calculates the future demand forecast value of taxis and the future demand forecast value of online car-hailing vehicles by the following formula:

[0081]

[0082]

[0083] Where: They represent the future demand forecast values ​​of taxis and online ride-hailing services respectively; denotes the taxi future demand matrix and the online car-hailing future demand matrix output by the b-th spatiotemporal block in the taxi component and the online car-hailing component, respectively; B denotes the number of spatiotemporal blocks; Respectively represent the connection of the taxi future demand matrix and the online car-hailing future demand matrix output by the taxi component and the online car-hailing component; and represents the learnable weight; σ(·) represents the ReLU activation function.

[0084] Preferably, when training the demand forecasting model, the model parameters of the demand forecasting model are optimized by the following loss function:

[0085]

[0086]

[0087]

[0088] Where: represents the loss function of the demand forecasting model; represents the mean square error of the taxi component; represents the mean square error of the online car-hailing component; They represent the future demand forecasts for taxis and online ride-hailing services respectively; Y TA , Y RS They represent the true value of future taxi demand and the true value of future online car-hailing demand respectively; σ1 and σ2 represent noise parameters.

[0089] The combined demand forecasting method based on taxis and online car-hailing vehicles in the present invention has the following beneficial effects:

[0090] The present invention regards the demand forecast of taxis and online-hailing cars as two coupled and related tasks, and extracts spatiotemporal features based on the historical demand matrix of taxis and the historical demand matrix of online-hailing cars, respectively, so that the temporal and spatial dependencies can be effectively extracted from the two traffic modes of taxis and online-hailing cars, that is, the intra-mode features of taxis and online-hailing cars can be effectively extracted; at the same time, the present invention generates corresponding spatiotemporal sharing information of taxis and online-hailing cars based on the historical demand matrix of taxis and the historical demand matrix of online-hailing cars, respectively, and then combines the intra-mode features of taxis and online-hailing cars to realize the transmission and exchange of temporal and spatial information between different modes, that is, the inter-mode features of taxis and online-hailing cars can be effectively extracted, and the fusion of inter-mode features and intra-mode features can be realized. Therefore, the present invention can extract the intra-mode features and inter-mode features of taxis and online-hailing cars, and can realize the fusion of intra-mode features and inter-mode features, and then realize the joint prediction of the demand of taxis and online-hailing cars, thereby improving the accuracy and comprehensiveness of traffic demand prediction, and providing a new idea for the joint prediction of the demand of multiple modes of transportation. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0092] Figure 1 The logical block diagram of the joint forecasting method for taxi and online car-hailing demand;

[0093] Figure 2 This is the network structure diagram of the demand forecasting model;

[0094] Figure 3 This is the working principle diagram of the time interleaving module;

[0095] Figure 4 This is the working principle diagram of the space interweaving module;

[0096] Figure 5 A diagram of the network structure of stacked spatiotemporal blocks. DETAILED DESCRIPTION

[0097] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions 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, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0098] It should be noted that similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. In the description of the present invention, it should be noted that the orientation or position relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inside", "outside", etc. is based on the orientation or position relationship shown in the drawings, or the orientation or position relationship in which the invention product is usually placed when used, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. In addition, the terms "first", "second", "third", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance. In addition, the terms "horizontal", "vertical", etc. do not mean that the components are required to be absolutely horizontal or suspended, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted. In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0099] The following is a further detailed description through specific implementation methods:

[0100] Example:

[0101] This embodiment discloses a joint demand prediction method based on taxis and online-hailing vehicles.

[0102] like Figure 1As shown in FIG. 1 , the joint demand forecasting method based on taxis and online car-hailing services includes:

[0103] S1: Obtain the corresponding taxi historical demand matrix and online car-hailing historical demand matrix;

[0104] S2: Input the taxi historical demand matrix and the online car-hailing historical demand matrix into the trained demand forecasting model, and output the corresponding taxi future demand forecast value and online car-hailing future demand forecast value;

[0105] The demand forecasting model is built based on interwoven neural networks;

[0106] The demand forecasting model first generates the corresponding taxi spatiotemporal sharing information and online car-hailing spatiotemporal sharing information based on the taxi historical demand matrix and the online car-hailing historical demand matrix combined with the corresponding adaptive adjacency matrix; then, the spatiotemporal features are extracted based on the taxi historical demand matrix and the corresponding adaptive adjacency matrix, and the corresponding taxi future demand matrix is ​​generated in combination with the online car-hailing spatiotemporal sharing information; at the same time, the spatiotemporal features are extracted based on the online car-hailing historical demand matrix and the corresponding adaptive adjacency matrix, and the corresponding online car-hailing future demand matrix is ​​generated in combination with the taxi spatiotemporal sharing information; finally, the taxi future demand forecast value and the online car-hailing future demand forecast value are generated based on the taxi future demand matrix and the online car-hailing future demand matrix;

[0107] S3: The future demand forecast value of taxis and the future demand forecast value of online ride-hailing vehicles are used as the joint demand forecast result of taxis and online ride-hailing vehicles.

[0108] It should be noted that the demand joint prediction method based on the present invention can be inserted into more other modes of transportation (such as subways, bicycles, and electric vehicles) to achieve higher accuracy than single prediction.

[0109] The present invention regards the demand forecast of taxis and online-hailing cars as two coupled and related tasks, and extracts spatiotemporal features based on the historical demand matrix of taxis and the historical demand matrix of online-hailing cars, respectively, so that the temporal and spatial dependencies can be effectively extracted from the two traffic modes of taxis and online-hailing cars, that is, the intra-mode features of taxis and online-hailing cars can be effectively extracted; at the same time, the present invention generates corresponding spatiotemporal sharing information of taxis and online-hailing cars based on the historical demand matrix of taxis and the historical demand matrix of online-hailing cars, respectively, and then combines the intra-mode features of taxis and online-hailing cars to realize the transmission and exchange of temporal and spatial information between different modes, that is, the inter-mode features of taxis and online-hailing cars can be effectively extracted, and the fusion of inter-mode features and intra-mode features can be realized. Therefore, the present invention can extract the intra-mode features and inter-mode features of taxis and online-hailing cars, and can realize the fusion of intra-mode features and inter-mode features, and then realize the joint prediction of the demand of taxis and online-hailing cars, thereby improving the accuracy and comprehensiveness of traffic demand prediction, and providing a new idea for the joint prediction of the demand of multiple modes of transportation.

[0110] In order to better illustrate the technical solution of the present invention, the following definitions are first made in this embodiment:

[0111] Definition 1 (Region): A city can be divided into multiple non-overlapping regions. Let R = {r1, r2, …, r N} represents the set of all N areas in the city. These areas can be regular or irregular. In this embodiment, the city area is obtained according to the postal code, so it is irregular.

[0112] Definition 2 (Time Interval): We divide the entire study time into a non-overlapping sequence with equal time intervals (Δt), for example, 30 minutes. Let T = {t1, t2, …, r L+1} represents a time series consisting of L time intervals in total. We also define the jth time interval as the interval from t j to j+1 Usually, the time interval is fixed and predefined according to the application scenario or device sampling frequency.

[0113] Definition 3 (Regional demand): Regional demand refers to the number of passengers who need to travel in a given region during an observation period. It can be expressed as i The number of passengers who take taxis or ride-hailing services in . Therefore, we have two types of regional demand, expressed as and where i and j represent the index of the region and time interval respectively.

[0114] Definition 4 (Regional Demand Matrix): We can use the demand matrix to represent the demand of N regions in the historical L time intervals. More specifically, the i-th row vector in the matrix (i.e., X(i,:)) records the demand of region r i The jth column vector in the matrix (i.e., X(:,j)) collects the demand of all N regions in the jth time interval. We use two notations, namely, X TA ∈R N×L and X RA ∈R N×L To distinguish the demand matrix of taxis and online ride-hailing services.

[0115] Definition 5 (Region adjacency matrix): Given a set of regions (i.e., {r1, r2, …, r N}), the region adjacency matrix A∈R N×N Used to describe the spatial relationship between regions. Typically, the adjacency matrix can be predefined or adaptive. The predefined ones are based on a measure of geographic / semantic proximity between regions, while the adaptive ones are based on the update of learnable parameters.

[0116] Definition 6 (Regional Demand Graph): The regional demand graph is defined as G = (V, A, X), where V is a set of nodes (i.e., N regions); A ∈ R N×N is the region adjacency matrix; X∈R N×L Refers to the regional demand matrix of all nodes in G within the historical L time interval.

[0117] Definition 7 (Temporal Convolution Operation): The temporal convolution operation aims to extract temporal features from historical demand sequences. Specifically, given a demand sequence The i-th region r i , and the core The size of is K, and the operation on the sequence at the jth position can be expressed as:

[0118]

[0119] where w s (1≤s≤K) represents the weight of the kernel, and d is a special parameter called the dilation factor, which is used to control the spacing between kernel points. When d=1, the operation is equivalent to a standard convolution, and when d>1 it is called a dilated convolution.

[0120] It is important to note that the temporal convolution is defined on the row vectors of the region demand matrix. The length of the new sequence after the temporal convolution becomes shorter, depending on the kernel size and the dilation factor.

[0121] Definition 8 (Spatial Convolution Operation): Spatial convolution is used to capture the spatial dependencies between different regions according to their demands. This operation requires an adjacency matrix A to specify the spatial relationships between regions and a demand matrix X to provide the spatial distribution of demands in all regions. In particular, different regions can be organized into graphs We use graph convolution to extract spatial dependencies.

[0122]

[0123] Where AX represents the inter-regional demand aggregation process based on the adjacency matrix; W represents a linear transformation, which is used to map the regional demand characteristics to a higher-dimensional space; σ(·) is an activation function, such as the Sigmoid function.

[0124] In this embodiment, the goal of the demand forecasting model (hereinafter referred to as TSIN) is to give the adaptive adjacency matrix A θ Initialized area map And the historical demand for taxis and online ride-hailing vehicles in the past L time intervals: X TA ∈R N×L and X RS ∈R N×L , predicting the future in the upcoming time interval requires all regions and Right now in is the function implemented by the neural network model. The goal is to determine the optimal function parameter Θ by minimizing the error between the estimated value and the true value * : in Represents the loss function.

[0125] Combination Figure 2 As shown, the demand forecasting model includes:

[0126] The input layer is used to obtain the historical demand matrix of taxis and the historical demand matrix of online ride-hailing vehicles;

[0127] A spatiotemporal interweaving component is used to generate corresponding taxi spatiotemporal sharing information and online car-hailing spatiotemporal sharing information according to the taxi historical demand matrix and the online car-hailing historical demand matrix in combination with the corresponding adaptive adjacency matrix;

[0128] The taxi component is used to extract spatiotemporal features based on the taxi historical demand matrix and the corresponding adaptive adjacency matrix, and integrate the spatiotemporal sharing information of online ride-hailing to generate the corresponding taxi future demand matrix;

[0129] The online car-hailing component is used to extract spatiotemporal features based on the online car-hailing historical demand matrix and the corresponding adaptive adjacency matrix, and integrate the taxi spatiotemporal sharing information to generate the corresponding online car-hailing future demand matrix;

[0130] The output layer is used to perform demand forecasting based on the taxi future demand matrix and the online car-hailing future demand matrix through a fully connected network, and generate and output the corresponding taxi future demand forecast value and online car-hailing future demand forecast value.

[0131] In the specific implementation process, the time-space interleaving component includes a time interleaving module and a space interleaving module;

[0132] Both the taxi component and the online car-hailing component include several spatiotemporal blocks (hereinafter referred to as ST blocks); each spatiotemporal block includes a corresponding temporal convolutional layer (hereinafter referred to as TCL) and a spatial convolutional layer (hereinafter referred to as SCL);

[0133] The time interleaving module is used to perform time convolution on the taxi historical demand matrix and the online car-hailing historical demand matrix respectively to generate corresponding taxi time sharing information and online car-hailing time sharing information;

[0134] For a single spatiotemporal block of a taxi component: extract temporal features based on the taxi historical demand matrix through a temporal convolutional layer, and integrate the online car-hailing time sharing information to generate the corresponding taxi demand temporal feature matrix;

[0135] For a single spatiotemporal block of the online car-hailing component: extract temporal features based on the online car-hailing historical demand matrix through the temporal convolutional layer, and integrate the taxi time sharing information to generate the corresponding online car-hailing demand time feature matrix;

[0136] The spatial interleaving module is used to extract spatial features based on the taxi demand time feature matrix and the online car-hailing demand time feature matrix combined with the corresponding adaptive adjacency matrix, and generate corresponding taxi spatiotemporal sharing information and online car-hailing spatiotemporal sharing information;

[0137] For a single spatiotemporal block of a taxi component: spatial features are extracted based on the taxi demand temporal feature matrix and the corresponding adaptive adjacency matrix through a spatial convolutional layer, and the spatiotemporal sharing information of online ride-hailing services is integrated to generate the corresponding taxi demand spatiotemporal feature matrix;

[0138] For a single spatiotemporal block of the online car-hailing component: spatial features are extracted based on the online car-hailing demand temporal feature matrix and the corresponding adaptive adjacency matrix through the spatial convolution layer, and the taxi spatiotemporal sharing information is integrated to generate the corresponding online car-hailing demand spatiotemporal feature matrix.

[0139] In the specific implementation, in order to capture the inter-modal temporal and spatial dependencies between taxi and ride-hailing demands, we propose a spatiotemporal interleaving component. The temporal interleaving component is used to connect two separate TCLs so that they can exchange valuable information in the temporal dependency model.

[0140] Combination Figure 3 As shown, the time interleaving module includes a first time convolution layer and a second time convolution layer for performing time convolution on the taxi historical demand matrix and the online car-hailing historical demand matrix respectively;

[0141] The temporal convolution formulas for the first temporal convolution layer and the second temporal convolution layer are as follows:

[0142]

[0143]

[0144] Where: They represent taxi time sharing information and online car-hailing time sharing information respectively; X TA , X RS They represent the historical demand matrix of taxis and the historical demand matrix of online ride-hailing cars respectively; TCL1 and TCL2 represent the first time convolution layer and the second time convolution layer respectively;

[0145] In order to avoid transmitting irrelevant information between different sequences, the time interleaving module filters the shared information through a self-gating mechanism combined with the following formula:

[0146]

[0147]

[0148] Where: They represent taxi time sharing information and online car-hailing time sharing information respectively; σ represents the Sigmod function; ⊙ represents the Hadamard product (element-wise product) operator.

[0149] The temporal convolution layer is built based on the dilated causal convolution, which can extract information from long sequences and is used to extract the temporal features of taxis and online ride-hailing vehicles.

[0150] Taking taxi as an example, for area r i , given its taxi demand sequence and kernel f Tt =[w0,w1,…,w K-1 ], applied to X TA The dilated causal convolution at the jth (1≤j≤L) position (i,:) can be expressed as: where K is the kernel size; d is the dilation factor that controls the jump distance. In order to capture more different patterns from the entire time series, we TA The L positions of (i,:) use F t Different kernels are used for convolution operations, which constitute our TCL: where X′ TA (i,:) is the new feature sequence obtained after time convolution, and the new length is L′=Ld×(K-1); F t is the number of channels (also called feature dimension in TCL). TCL can perform the following operations on the demand sequences of N regions simultaneously: where X′ AA It is the new demand matrix for taxis and also the final output of TCL.

[0151] Therefore, for a single spatiotemporal block of the taxi component: the taxi historical demand matrix is ​​used as the input of the temporal convolution layer and temporally convolved to generate the corresponding taxi demand temporal feature matrix; then the taxi demand temporal feature matrix and the online car-hailing time sharing information are fused in the last channel of the time dimension to obtain the final taxi demand temporal feature matrix;

[0152] The formula is described as:

[0153]

[0154]

[0155] Where: X′ TA represents the taxi demand time characteristic matrix; X TA represents the taxi historical demand matrix; F t represents the number of channels; L′ represents the length of the time dimension; X′ TA (:,L′,:), They represent the taxi demand time feature matrix and the last channel of the online car-hailing time sharing information in the time dimension respectively;

[0156] For a single spatiotemporal block of the online car-hailing component: the online car-hailing historical demand matrix is ​​used as the input of the temporal convolution layer and temporally convolved to generate the corresponding online car-hailing demand time feature matrix; then the online car-hailing demand time feature matrix and the taxi time sharing information are fused in the last channel of the time dimension to obtain the final online car-hailing demand time feature matrix;

[0157] The formula is described as:

[0158]

[0159]

[0160] Where: X′ RS represents the time characteristic matrix of online car-hailing demand; X RS represents the historical demand matrix of online ride-hailing; F t represents the number of channels; L′ represents the length of the time dimension; X′ RS (:,L′,:), They respectively represent the last channel of the online car-hailing demand time feature matrix and the taxi time sharing information in the time dimension.

[0161] Combination Figure 4 As shown, the spatial interleaving module includes a first spatial convolution layer and a second spatial convolution layer for performing spatial convolution on the taxi demand time feature matrix and the online car-hailing demand time feature matrix respectively;

[0162] The spatial convolution formulas of the first spatial convolution layer and the second spatial convolution layer are as follows:

[0163]

[0164]

[0165]

[0166]

[0167] Where: Respectively represent the time-space sharing information of taxis and online car-hailing; A RT , A TR Represent the adaptive adjacency matrix between taxi and online car-hailing demands, where A RT represents the relationship between the taxi demand in one area and the online car-hailing demand in other areas. TR represents the relationship between the online car-hailing demand in one area and the taxi demand in other areas; X′ TA , X′ RS They represent the taxi demand time characteristic matrix and the online car-hailing demand time characteristic matrix respectively; represents two feature transformation matrices; E Tq , represents the regional characteristics set for taxi demand; E Rq , represents the regional characteristics set for online ride-hailing demand; N represents the number of regions;

[0168] The spatial convolutional layer has two tasks, graph learning (GL) and graph convolution (GC).

[0169] Taking taxis as an example, in GL, we initialize two regional features for taxi demand where d e is the dimension of the feature. The calculation formula of the adaptive adjacency matrix of taxi demand between N regions is: Where E Tq ,E Tk is composed of learnable parameters. Tq ·E Tk It is used to calculate the similarity of feature vectors between different regions, which can be regarded as the weight of spatial dependence. The activation function ReLU is used to remove small weights, and the Softmax function is used to normalize the matrix. Adaptive adjacency matrix A TT It can be automatically updated according to the error feedback of the prediction task during the training phase. TT Finally, we use graph convolution (GC) to model the intra-mode spatial dependencies between N regions. Again, GC only focuses on a single traffic mode. Specifically, the graph convolution layer converts A TT and X′ TA As input, a new demand feature X″ is generated for each region through aggregation and transformation operations. TA , in the form of: in, is the projection used for feature transformation; σ(·) represents the GLU activation function, F s is the output dimension of the graph convolutional layer. Similarly, with the same network structure as taxi demand, we also constructed two regional features for online car-hailing demand And the adaptive adjacency matrix is ​​expressed as: The corresponding graph convolutional layer is defined as: in is the matrix used for linear transformation.

[0170] Therefore, for a single spatiotemporal block of the taxi component: the taxi demand temporal feature matrix is ​​used as the input of the spatial convolution layer and combined with the corresponding adaptive adjacency matrix for spatial convolution to generate the corresponding taxi demand spatiotemporal feature matrix; then the taxi demand spatiotemporal feature matrix is ​​fused with the online car-hailing spatiotemporal sharing information to obtain the final taxi demand spatiotemporal feature matrix;

[0171] The formula is described as:

[0172]

[0173]

[0174] The graph learning sublayer is used to automatically discover the dependencies between different regions and generate the corresponding adaptive adjacency matrix.

[0175] The formula is described as:

[0176]

[0177] Where: X″ TA represents the spatiotemporal characteristic matrix of taxi demand; X′ TA A represents the taxi demand time characteristic matrix; TT An adaptive adjacency matrix representing taxi demand between N regions; represents the projection used for feature transformation; σ(·) represents the GLU activation function; F s Represents the output dimension of the graph convolutional layer; represents the time-space sharing information of online car-hailing; E Tq , represents the regional characteristics set for taxi demand; N represents the number of regions;

[0178] For a single spatiotemporal block of the online car-hailing component: the online car-hailing demand temporal feature matrix is ​​used as the input of the spatial convolution layer and combined with the corresponding adaptive adjacency matrix for spatial convolution to generate the corresponding online car-hailing demand spatiotemporal feature matrix; then the online car-hailing demand spatiotemporal feature matrix is ​​fused with the taxi spatiotemporal sharing information to obtain the final online car-hailing demand spatiotemporal feature matrix;

[0179] The formula is described as:

[0180]

[0181]

[0182] The graph learning sublayer is used to automatically discover the dependencies between different regions and generate the corresponding adaptive adjacency matrix.

[0183] The formula is described as:

[0184]

[0185] Where: X ′ R ″ S represents the spatiotemporal characteristic matrix of online car-hailing demand; X ′ RS represents the time characteristic matrix of online car-hailing demand; W RR An adaptive adjacency matrix representing the demand for online ride-hailing services between N regions; represents the projection used for feature transformation; σ(·) represents the GLU activation function; F s Represents the output dimension of the graph convolutional layer; represents the taxi space-time sharing information; E Rq , Represents the regional characteristics set for online ride-hailing demand; N represents the number of regions.

[0186] In the present invention, the demand forecasting model regards the demand forecasting of taxis and online-hailing cars as two coupled and related tasks, and can extract the time characteristics and spatial characteristics of the historical demand matrix of taxis and the historical demand matrix of online-hailing cars in turn, and then can effectively extract the temporal and spatial dependencies from the corresponding traffic modes, that is, it can ensure the extraction effect of the characteristics within the taxi and online-hailing modes. At the same time, the present invention generates time sharing information and space sharing information based on the historical demand matrix of taxis and the historical demand matrix of online-hailing cars, and fuses the time sharing information and space sharing information at the corresponding stages of extracting time characteristics and space characteristics, so that the time and space information transmission and exchange between different modes can be better realized, and then the inter-mode characteristics of taxis and online-hailing cars can be effectively extracted and integrated with the intra-mode characteristics, so as to further improve the accuracy and comprehensiveness of the joint prediction of taxi and online-hailing car demand.

[0187] Combination Figure 5 As shown, TCL and SCL together form an ST block to capture spatiotemporal dependencies simultaneously. In order to avoid the problem of gradient vanishing when stacking blocks, a residual connection structure is set in the spatiotemporal block;

[0188] For a single spatiotemporal block of the taxi component: the taxi historical demand matrix and the corresponding taxi demand spatiotemporal feature matrix are fused to generate the taxi demand matrix of the spatiotemporal block;

[0189] The formula is described as:

[0190]

[0191] Where: represents the taxi demand matrix of the b+1th spatiotemporal block in the taxi component; represents the taxi historical demand matrix of the b-th spatiotemporal block input in the taxi component; represents the temporal convolution of the b-th spatiotemporal block in the taxi component; represents the spatial convolution of the b-th spatiotemporal block in the taxi component;

[0192] For a single spatiotemporal block of the online ride-hailing component: the online ride-hailing historical demand matrix and the corresponding online ride-hailing demand spatiotemporal feature matrix are merged to generate the online ride-hailing demand matrix of the spatiotemporal block;

[0193] The formula is described as:

[0194]

[0195] Where: represents the online car-hailing demand matrix of the b+1th spatiotemporal block in the online car-hailing component; Represents the historical demand matrix of online ride-hailing input in the b-th spatiotemporal block in the online ride-hailing component; represents the temporal convolution of the b-th spatiotemporal block in the online car-hailing component; Represents the spatial convolution of the b-th spatiotemporal block in the ride-hailing component.

[0196] Combination Figure 5 As shown, a jump connection structure is set between the space-time blocks;

[0197] The taxi component and the online car-hailing component respectively connect the taxi demand matrices and online car-hailing demand matrices of all their time-space blocks together to generate the corresponding taxi future demand matrices and online car-hailing future demand matrices.

[0198] The present invention connects the outputs of each space-time block to generate a corresponding future demand matrix through the residual connection structure in the space-time block and the jump connection structure between the space-time blocks, so that the space-time dependency of taxis and online car-hailing vehicles can be captured simultaneously, and the problem of gradient vanishing when the space-time blocks are stacked can be avoided. The fusion effect of inter-mode features and intra-mode features can be ensured, thereby further improving the accuracy of the joint prediction of taxi and online car-hailing demand.

[0199] In the specific implementation process, the output layer calculates the future demand forecast values ​​of taxis and online car-hailing vehicles through the following formula:

[0200]

[0201]

[0202] Where: They represent the future demand forecast values ​​of taxis and online ride-hailing services respectively; denotes the taxi future demand matrix and the online car-hailing future demand matrix output by the b-th spatiotemporal block in the taxi component and the online car-hailing component, respectively; B denotes the number of spatiotemporal blocks; Respectively represent the connection of the taxi future demand matrix and the online car-hailing future demand matrix output by the taxi component and the online car-hailing component; and represents the learnable weight; σ(·) represents the ReLU activation function.

[0203] When training the demand forecasting model, the model parameters of the demand forecasting model are optimized through the following loss function:

[0204]

[0205]

[0206]

[0207] Where: represents the loss function of the demand forecasting model; represents the mean square error of the taxi component; represents the mean square error of the online car-hailing component; They represent the future demand forecasts for taxis and online ride-hailing services respectively; Y TA , Y RS They represent the true value of future taxi demand and the true value of future online car-hailing demand respectively; σ1 and σ2 represent noise parameters, which are used to balance the loss of specific tasks during training and can be updated by back propagation.

[0208] The present invention uses mean square error as the loss function of the demand forecasting model and sets corresponding noise parameters, so as to ensure the effectiveness and effect of the demand forecasting model training, thereby improving the performance of the demand forecasting model, thereby further improving the accuracy of the joint forecast of taxi and online car-hailing demand.

[0209] In order to better illustrate the advantages of the technical solution of the present invention, this embodiment also discloses the following experiment.

[0210] 1. Dataset

[0211] This experiment was conducted on four real datasets generated by taxis and online ride-hailing vehicles in two representative cities in the United States (New York NYC and Chicago CHI). For each dataset, we set the time interval to 30 minutes, and the training set, validation set, and test set were divided into 7:1:2. We used Z-Score to normalize the input features and used the historical demand of the last 12 time periods to predict the demand of the next time period. The dataset is described in detail as follows:

[0212] Table 1 Dataset description

[0213]

[0214] 2. Experimental Setup

[0215] This experiment uses Pytorch to implement our TSIN model on a Linux workstation (GPU: GeForce RTX 2080Ti). The maximum number of training iterations is set to 500. The learning rate is set to 0.001 and the batch size is 64. We use the Adam optimizer to minimize the loss function.

[0216] 3. Experimental methods

[0217] This experiment will be compared with 11 baselines. For fair comparison, the input of all methods on both datasets is the same, that is, the historical demand for the past L time intervals. For hyperparameters, we choose the default values ​​according to their recommendations.

[0218] Comparison criteria. To evaluate the performance of all methods, we use three indicators: mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE). They are defined as follows:

[0219]

[0220] where y and are the true value and the predicted value respectively. MAE is a widely used measure of absolute error; MAPE considers the ratio of absolute errors relative to the true value, but it penalizes smaller true values ​​more; RMSE is more sensitive to outliers. Therefore, the combination of the three indicators can more comprehensively evaluate the performance of these methods.

[0221] This experiment shows the comparison results between our model and other 11 state-of-the-art baselines.

[0222] As shown in Table 2. Our method TSIN achieves the best performance and the lowest error among all baselines on the New York dataset. The MAE and RMSE values ​​of TSIN on the Chicago taxi and ride-hailing datasets are the lowest in their corresponding columns. This shows that our model achieves the lowest absolute error. However, the MAPE values ​​are relatively high, especially on the taxi dataset. The reason may be that MAPE is sensitive to the actual demand, and the taxi demand in Chicago decreased as the taxi market declined at that time.

[0223] Table 2. Comparison of performance of different baselines on New York and Chicago datasets

[0224]

[0225] 4. Computational efficiency

[0226] This experiment evaluates the computational efficiency based on the NYC dataset from three aspects: the number of model parameters, training time, and inference time. Our model and all neural network-based baselines are evaluated under the same conditions for fair comparison. The results are shown in Table 3. It can be seen that TSIN has a moderate model complexity, and the training and inference time are relatively short, that is, the prediction efficiency is higher.

[0227] Table 3 Results of calculating efficiency

[0228]

[0229] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. A joint forecasting method based on taxi and online car-hailing demand, characterized by: include: S1: Obtain the corresponding taxi historical demand matrix and online car-hailing historical demand matrix; S2: Input the taxi historical demand matrix and the online car-hailing historical demand matrix into the trained demand forecasting model, and output the corresponding taxi future demand forecast value and online car-hailing future demand forecast value; The demand forecasting model first generates the corresponding taxi spatiotemporal sharing information and online car-hailing spatiotemporal sharing information based on the taxi historical demand matrix and the online car-hailing historical demand matrix combined with the corresponding adaptive adjacency matrix; then, the spatiotemporal features are extracted based on the taxi historical demand matrix and the corresponding adaptive adjacency matrix, and the corresponding taxi future demand matrix is ​​generated in combination with the online car-hailing spatiotemporal sharing information; at the same time, the spatiotemporal features are extracted based on the online car-hailing historical demand matrix and the corresponding adaptive adjacency matrix, and the corresponding online car-hailing future demand matrix is ​​generated in combination with the taxi spatiotemporal sharing information; finally, the taxi future demand forecast value and the online car-hailing future demand forecast value are generated based on the taxi future demand matrix and the online car-hailing future demand matrix; Demand forecasting models include: The input layer is used to obtain the historical demand matrix of taxis and the historical demand matrix of online ride-hailing vehicles; A spatiotemporal interweaving component is used to generate corresponding taxi spatiotemporal sharing information and online car-hailing spatiotemporal sharing information according to the taxi historical demand matrix and the online car-hailing historical demand matrix in combination with the corresponding adaptive adjacency matrix; The taxi component is used to extract spatiotemporal features based on the taxi historical demand matrix and the corresponding adaptive adjacency matrix, and integrate the spatiotemporal sharing information of online ride-hailing to generate the corresponding taxi future demand matrix; The online car-hailing component is used to extract spatiotemporal features based on the online car-hailing historical demand matrix and the corresponding adaptive adjacency matrix, and integrate the taxi spatiotemporal sharing information to generate the corresponding online car-hailing future demand matrix; The output layer is used to perform demand forecasting based on the taxi future demand matrix and the online car-hailing future demand matrix through a fully connected network, and generate and output the corresponding taxi future demand forecast value and online car-hailing future demand forecast value; The space-time interleaving component includes a time interleaving module and a space interleaving module; Both the taxi component and the online car-hailing component include several spatiotemporal blocks; each spatiotemporal block includes corresponding temporal convolutional layers and spatial convolutional layers; The time interleaving module is used to perform time convolution on the taxi historical demand matrix and the online car-hailing historical demand matrix respectively to generate corresponding taxi time sharing information and online car-hailing time sharing information; For a single spatiotemporal block of a taxi component: extract temporal features based on the taxi historical demand matrix through a temporal convolutional layer, and integrate the online car-hailing time sharing information to generate the corresponding taxi demand temporal feature matrix; For a single spatiotemporal block of the online car-hailing component: extract temporal features based on the online car-hailing historical demand matrix through the temporal convolutional layer, and integrate the taxi time sharing information to generate the corresponding online car-hailing demand time feature matrix; The spatial interleaving module is used to extract spatial features based on the taxi demand time feature matrix and the online car-hailing demand time feature matrix combined with the corresponding adaptive adjacency matrix, and generate corresponding taxi spatiotemporal sharing information and online car-hailing spatiotemporal sharing information; For a single spatiotemporal block of a taxi component: spatial features are extracted based on the taxi demand temporal feature matrix and the corresponding adaptive adjacency matrix through a spatial convolutional layer, and the spatiotemporal sharing information of online ride-hailing services is integrated to generate the corresponding taxi demand spatiotemporal feature matrix; For a single spatiotemporal block of the online car-hailing component: spatial features are extracted based on the online car-hailing demand temporal feature matrix and the corresponding adaptive adjacency matrix through the spatial convolution layer, and the taxi spatiotemporal sharing information is integrated to generate the corresponding online car-hailing demand spatiotemporal feature matrix; The time interleaving module includes a first time convolution layer and a second time convolution layer for performing time convolution on the taxi historical demand matrix and the online car-hailing historical demand matrix respectively; The temporal convolution formulas for the first temporal convolution layer and the second temporal convolution layer are as follows: Where: They represent taxi time sharing information and online car-hailing time sharing information respectively; X TA , X RS They represent the historical demand matrix of taxis and the historical demand matrix of online ride-hailing cars respectively; TCL1 and TCL2 represent the first time convolution layer and the second time convolution layer respectively; For a single spatiotemporal block of the taxi component: the taxi historical demand matrix is ​​used as the input of the temporal convolution layer and temporally convolved to generate the corresponding taxi demand temporal feature matrix; then the taxi demand temporal feature matrix and the online car-hailing time sharing information are fused in the last channel of the time dimension to obtain the final taxi demand temporal feature matrix; The formula is described as: Where: X′ TA represents the taxi demand time characteristic matrix; X TA represents the taxi historical demand matrix; F t represents the number of channels; L′ represents the length of the time dimension; X′ TA (:,L′,:), They represent the taxi demand time feature matrix and the last channel of the online car-hailing time sharing information in the time dimension respectively; For a single spatiotemporal block of the online car-hailing component: the online car-hailing historical demand matrix is ​​used as the input of the temporal convolution layer and temporally convolved to generate the corresponding online car-hailing demand time feature matrix; then the online car-hailing demand time feature matrix and the taxi time sharing information are fused in the last channel of the time dimension to obtain the final online car-hailing demand time feature matrix; The formula is described as: Where: X′ RS represents the time characteristic matrix of online car-hailing demand; X RS represents the historical demand matrix of online ride-hailing; F t represents the number of channels; L′ represents the length of the time dimension; X′ RS (:,L′,:), They represent the last channel of the online car-hailing demand time feature matrix and taxi time sharing information in the time dimension respectively; The spatial interleaving module includes a first spatial convolution layer and a second spatial convolution layer for performing spatial convolution on the taxi demand time feature matrix and the online car-hailing demand time feature matrix respectively; The spatial convolution formulas of the first spatial convolution layer and the second spatial convolution layer are as follows: Where: They represent the time-space sharing information of taxis and the time-space sharing information of online car-hailing services respectively; A RT , A TR Represent the adaptive adjacency matrix between taxi and online car-hailing demands, where A RT represents the relationship between the taxi demand in one area and the online car-hailing demand in other areas. TR represents the relationship between the online car-hailing demand in one area and the taxi demand in other areas; X′ TA , X′ RS They represent the taxi demand time characteristic matrix and the online car-hailing demand time characteristic matrix respectively; represents two feature transformation matrices; E Tq , represents the regional characteristics set for taxi demand; E Rq , represents the regional characteristics set for online ride-hailing demand; N represents the number of regions; For a single spatiotemporal block of a taxi component: the taxi demand temporal feature matrix is ​​used as the input of the spatial convolution layer and combined with the corresponding adaptive adjacency matrix for spatial convolution to generate the corresponding taxi demand spatiotemporal feature matrix; then the taxi demand spatiotemporal feature matrix is ​​fused with the online car-hailing spatiotemporal sharing information to obtain the final taxi demand spatiotemporal feature matrix; The formula is described as: The graph learning sublayer is used to automatically discover the dependencies between different regions and generate the corresponding adaptive adjacency matrix. The formula is described as: Where: X″ TA represents the spatiotemporal characteristic matrix of taxi demand; X′ TA A represents the taxi demand time characteristic matrix; TT An adaptive adjacency matrix representing taxi demand between N regions; represents the projection used for feature transformation; σ(·) represents the GLU activation function; F s Represents the output dimension of the graph convolutional layer; represents the time-space sharing information of online car-hailing; E Tq , represents the regional characteristics set for taxi demand; N represents the number of regions; For a single spatiotemporal block of the online car-hailing component: the online car-hailing demand temporal feature matrix is ​​used as the input of the spatial convolution layer and combined with the corresponding adaptive adjacency matrix for spatial convolution to generate the corresponding online car-hailing demand spatiotemporal feature matrix; then the online car-hailing demand spatiotemporal feature matrix is ​​fused with the taxi spatiotemporal sharing information to obtain the final online car-hailing demand spatiotemporal feature matrix; The formula is described as: The graph learning sublayer is used to automatically discover the dependencies between different regions and generate the corresponding adaptive adjacency matrix. The formula is described as: Where: X″ RS represents the spatiotemporal characteristic matrix of online car-hailing demand; X′ RS A represents the time characteristic matrix of online car-hailing demand; RR An adaptive adjacency matrix representing the demand for online ride-hailing services between N regions; represents the projection used for feature transformation; σ(·) represents the GLU activation function; F s Represents the output dimension of the graph convolutional layer; represents the taxi space-time sharing information; E Rq , represents the regional characteristics set for online ride-hailing demand; N represents the number of regions; A residual connection structure is set in the spatiotemporal block; For a single spatiotemporal block of the taxi component: the taxi historical demand matrix and the corresponding taxi demand spatiotemporal feature matrix are fused to generate the taxi demand matrix of the spatiotemporal block; The formula is described as: Where: represents the taxi demand matrix of the b+1th spatiotemporal block in the taxi component; represents the taxi historical demand matrix of the b-th spatiotemporal block input in the taxi component; represents the temporal convolution of the b-th spatiotemporal block in the taxi component; represents the spatial convolution of the b-th spatiotemporal block in the taxi component; For a single spatiotemporal block of the online ride-hailing component: the online ride-hailing historical demand matrix and the corresponding online ride-hailing demand spatiotemporal feature matrix are merged to generate the online ride-hailing demand matrix of the spatiotemporal block; The formula is described as: Where: represents the online car-hailing demand matrix of the b+1th spatiotemporal block in the online car-hailing component; Represents the historical demand matrix of online ride-hailing input in the b-th spatiotemporal block in the online ride-hailing component; represents the temporal convolution of the b-th spatiotemporal block in the online car-hailing component; represents the spatial convolution of the b-th spatiotemporal block in the online car-hailing component; A jump connection structure is set between the space-time blocks; The taxi component and the online car-hailing component respectively connect the taxi demand matrix and the online car-hailing demand matrix of all their time-space blocks to generate the corresponding taxi future demand matrix and the online car-hailing future demand matrix; The output layer calculates the future demand forecast values ​​of taxis and online car-hailing vehicles through the following formula: Where: They represent the future demand forecast values ​​of taxis and online ride-hailing services respectively; denotes the taxi future demand matrix and the online car-hailing future demand matrix output by the b-th spatiotemporal block in the taxi component and the online car-hailing component, respectively; B denotes the number of spatiotemporal blocks; Respectively represent connecting the taxi future demand matrix and the online car-hailing future demand matrix output by the taxi component and the online car-hailing component; and represents the learnable weight; σ(·) represents the ReLU activation function; S3: The future demand forecast values ​​of taxis and online ride-hailing vehicles are used as the joint demand forecast results of taxis and online ride-hailing vehicles, and ideas are provided for the joint demand forecast of various modes of transportation based on the joint demand forecast results.

2. The method for joint forecasting of taxi and online car-hailing demand according to claim 1, characterized in that: The time interleaving module filters the shared information through the self-gating mechanism combined with the following formula: Where: They represent taxi time sharing information and online car-hailing time sharing information respectively; σ represents the Sigmod function; ⊙ represents the element-wise product operator.

3. The method for joint forecasting of taxi and online car-hailing demand according to claim 1, characterized in that: When training the demand forecasting model, the model parameters of the demand forecasting model are optimized through the following loss function: Where: represents the loss function of the demand forecasting model; represents the mean square error of the taxi component; represents the mean square error of the online car-hailing component; They represent the future demand forecasts for taxis and online ride-hailing services respectively; Y Ta , Y RS They represent the true value of future taxi demand and the true value of future online car-hailing demand respectively; σ1 and σ2 represent noise parameters.

Images