A multimodal online ride-hailing demand uncertainty prediction method

Abstract

The application provides a multimode online car-hailing demand uncertainty prediction method. The method comprises the following steps: constructing a hybrid expert network structure composed of a dedicated expert module and a shared expert module, obtaining historical origin-destination (OD) demand data of a city multimode online car-hailing platform, and constructing an OD three-dimensional tensor comprising multiple travel modes; using the OD three-dimensional tensor of the multiple travel modes as training data, training the hybrid expert network structure using the training data, and obtaining a trained hybrid expert network structure; inputting the OD three-dimensional tensor of the multiple travel modes in a period of historical time into the trained hybrid expert network structure, and predicting the OD three-dimensional tensor of the multiple travel modes in a future time slice. Through the learning of the space-time dependence relationship and the operation dependence relationship, the accuracy of the carpooling demand prediction is improved, which provides important support for the platform to more reasonably allocate vehicle resources, meet the travel demand of users, and improve the service quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic demand, in particular to a multi-mode online car-hailing demand uncertainty prediction method. BACKGROUND

[0002] In recent years, with the development of the Internet and mobile terminals, a large number of order data recorded by online travel platforms provides rich training data for prediction models, and massive data enables deep learning methods to be widely used in carpool demand prediction. In terms of data sources, people have gradually focused on fusing multi-source data and multi-dimensional features, and by introducing meteorological, building environment, and interest point and other external factors, a more complex spatiotemporal prediction model is constructed.

[0003] One of the online car-hailing demand prediction methods in the prior art includes an STZINB-GNN model. The model is a probability prediction model combined with a zero-inflated negative binomial (ZINB) distribution, which realizes the online car-hailing demand prediction in a city. This model combines a zero-inflated statistical model and has the ability to quantify uncertainty.

[0004] The disadvantages of the online car-hailing demand prediction method in the prior art include that the online car-hailing demand prediction model in the prior art focuses on a single service mode (such as a private car), and cannot capture the interaction between multiple travel modes in a city. Especially in the city carpooling (carpooling) service, there is obvious sparsity, zero inflation, and temporal heterogeneity, and existing deep learning models are difficult to accurately model and predict. At the same time, most existing methods focus on "point prediction" of the average demand, lack effective quantification of prediction uncertainty, and are difficult to support robust scheduling and risk perception demand of the platform. SUMMARY

[0005] The present application provides a multi-mode online car-hailing demand uncertainty prediction method to accurately predict the demand for online car-hailing.

[0006] In order to achieve the above purpose, the present application adopts the following technical solutions.

[0007] A multi-mode online car-hailing demand uncertainty prediction method, comprising:

[0008] A hybrid expert network structure composed of exclusive expert modules and shared expert modules is constructed, and each exclusive expert module and shared expert module includes a graph convolution network GCN, a graph structure learning module GSL, and a time convolution network TCN;

[0009] The historical origin-destination (OD) demand data of a multi-mode online car-hailing platform in a city is obtained, and an OD three-dimensional tensor including multiple travel modes is constructed;

[0010] The OD three-dimensional tensor of multiple travel modes is taken as training data, and the hybrid expert network structure is trained by using the training data to obtain a trained hybrid expert network structure.

[0011] The OD three-dimensional tensor of multiple travel modes in a historical time period is input into the trained hybrid expert network structure, and the trained hybrid expert network structure outputs the predicted OD three-dimensional tensor of multiple travel modes in a future time slice.

[0012] Preferably, the hybrid expert network structure composed of exclusive expert modules and shared expert modules is constructed, and each exclusive expert module and shared expert module includes a graph convolution network (GCN), a graph structure learning (GSL) module, and a time convolution network (TCN), which includes:

[0013] A time-aware hybrid expert network composed of exclusive expert modules, shared expert modules, a gating module, and a quantile prediction module is constructed, and each exclusive expert module and shared expert module internally includes a space-time feature extraction submodule composed of a GCN and a TCN. A graph structure learning (GSL) mechanism is embedded in the GCN, and an adaptive adjacency matrix is reconstructed through singular value decomposition.

[0014] The shared expert module is configured to accept joint inputs of all travel modes and learn shared knowledge among modes. The exclusive expert module is configured to receive only inputs of a specific travel mode and model the specific travel mode.

[0015] The gating module is configured to dynamically activate different expert module combinations according to a current spatio-temporal context by using a time-aware router, weight and fuse output data from different expert modules, and transmit the weighted and fused data to an enhanced quantile prediction module.

[0016] The enhanced quantile prediction module is configured to generate non-crossing multi-quantile prediction values satisfying monotonicity based on the weighted and fused output data and a median regression result by using a residual construction mechanism, and derive prediction interval results at each confidence level.

[0017] Preferably, the historical origin-destination (OD) demand data of a city multi-modal ride-hailing platform is obtained, and an OD three-dimensional tensor including multiple travel modes is constructed, which includes:

[0018] The historical OD sequence of each travel mode is obtained, a multi-layer perception (MLP) is used to normalize and independently embed and encode the historical OD sequence of each travel mode to obtain a hidden layer representation of the MLP, and an OD three-dimensional tensor of each travel mode is constructed. The expression of the OD three-dimensional tensor is: T inN represents the length of the historical time slice of the input hybrid expert network, N represents the number of OD pairs, and m represents the number of travel mode types, and the travel modes include express cars, carpooling and taxis.

[0019] Preferably, the OD three-dimensional tensor of multiple travel modes is used as training data, and the training data is used to train the hybrid expert network structure to obtain a trained hybrid expert network structure, comprising:

[0020] The OD three-dimensional tensor of multiple travel modes As training data, after the training data is input into the hybrid expert network structure, it is first mapped through a linear layer to obtain an initial hidden layer code, which is transmitted to a dedicated expert branch and a shared expert module. The dedicated expert branch extracts and calculates the spatio-temporal features of the hidden layer code, and outputs a dedicated expert spatio-temporal code. The dedicated expert spatio-temporal code is then input into a gating module. The shared expert module first determines the active shared expert subset according to the time type to which the data belongs, and the active shared expert subset extracts and calculates the spatio-temporal features of the hidden layer code respectively, and finally gives the respective shared spatio-temporal code. The final shared spatio-temporal code is input into the gating module after being packed;

[0021] The gating module adaptively adjusts the weight occupied by different shared spatio-temporal codes according to the spatio-temporal context of the dedicated code, and finally fuses the shared spatio-temporal code with the dedicated expert spatio-temporal code to obtain a fused spatio-temporal code. The spatio-temporal context includes but is not limited to workday and weekend, morning and evening peak and flat peak scenario labels;

[0022] The fused spatio-temporal code is input into an enhanced quantile prediction module, and the enhanced quantile prediction module is composed of a group of parallel MLP+ReLU structures. Each layer predicts a residual term δ k Each quantile value is generated by recursively adding a non-negative offset based on the median prediction result The enhanced quantile prediction module constructs multiple upper and lower quantile prediction values by recursively accumulating non-negative offsets based on the median regression value, and outputs the OD prediction demand value of each travel mode under multiple confidence intervals.

[0023] The training loss function of the hybrid expert network structure is Huber smooth quantile loss. The value of the training loss function is used for back propagation, and the parameters of the hybrid expert network structure are iteratively updated. The training of the hybrid expert network structure is terminated until the change range of the value of the training loss function is less than a set threshold range, or a predetermined training termination condition is reached. The trained hybrid expert network structure is obtained.

[0024] Preferably, the method further comprises: ​

[0025] The shared expert module is used to accept the OD 3D tensor of all travel modes. Knowledge sharing between learning modes, sharing TCN in the expert module in T in Performing multi-layer temporal convolution operations in the dimension yields the temporal hidden code H. t The GCN in the shared expert module combines graph structure learning techniques with multi-layer graph convolution operations in the N-dimensional dimension to obtain the spatial hiding code H. s H t and H s At T respectively in The dimensions and N-dimensionality are processed using a multilayer perceptron, and then the Hadamard product of the processed results is processed again using a multilayer perceptron to obtain the shared spatiotemporal code of the shared expert.

[0026] A dedicated expert module is used to receive OD 3D tensors only for specific travel modes. Modeling specific travel patterns, the TCN in the dedicated expert module in T in Performing multi-layer temporal convolution operations in the dimension yields the temporal hidden code H. t The GCN in the dedicated expert module combines graph structure learning techniques with multi-layer graph convolution operations in the N-dimensional dimension to obtain the spatial hidden code H. s H t and H s At T respectively in The dimensions and N-dimensionality are processed using a multilayer perceptron, and then the Hadamard product of the processed results is processed again using a multilayer perceptron to obtain the dedicated expert spatiotemporal code.

[0027] Preferably, the step of inputting the OD three-dimensional tensors of multiple travel models over a historical period into the trained hybrid expert network structure, and the training hybrid expert network structure outputting the predicted OD three-dimensional tensors of multiple travel modes in future time slices, includes:

[0028] Obtain the OD 3D tensor of multiple travel models over a historical period. T in Let N represent the length of a time slice within a historical period, N represent the number of OD pairs, and M represent the number of travel mode types. The 3D tensor X of the ODs of these multiple travel models is input into a trained hybrid expert network structure. The output of the trained hybrid expert network structure predicts one of the future time slices, representing the set of 3D tensors of the ODs of the multiple travel models. in T outThe time slice length representing the output result of the hybrid expert network, N represents the number of OD pairs, M represents the number of travel mode types, tau is a quantile point, and T is a predefined quantile point set.

[0029] It can be seen from the technical solutions provided by the above-mentioned embodiments of the application that the application aims at short-term demand prediction of carpooling travel of a large range of OD (Origin-Destination) scales, improves the accuracy of carpooling demand prediction through learning of space-time dependence and operation dependence, and provides important support for platforms to more reasonably allocate vehicle resources, meet user travel demand, and improve service quality.

[0030] Additional aspects and advantages of the application will be described in the following description, which will become apparent from the following description, or will be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0032] Figure 1 A processing flow chart of a multi-mode online car-hailing demand uncertainty prediction method provided by the embodiment of the application is provided.

[0033] Figure 2 A module structure diagram of a time-aware hybrid expert network structure provided by the embodiment of the application is provided.

[0034] Figure 3 A structure diagram of a sliding window of a TCN in a space-time feature extraction sub-module provided by the embodiment of the application is provided.

[0035] Figure 4 An implementation principle diagram of an enhanced quantile prediction module provided by the embodiment of the application is provided. DETAILED DESCRIPTION

[0036] The embodiments of the application are described in detail below, and examples of the embodiments are shown in the drawings, wherein the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are only used to explain the application, and cannot be interpreted as a limitation on the application.

[0037] It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items. It will be understood by those within the art that, in general, terms used herein, and especially

[0038] Those skilled in the art will appreciate that unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the specification and relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0039] For the purpose of promoting an understanding of the principles of the present application, reference will now be made to the embodiments illustrated in the drawings, and specific language will be used to describe the same. It will, nevertheless, be understood that no limitation of the scope of the application is intended by illustrative discussion of the embodiments.

[0040] The method of the application firstly constructs a multi-mode OD three-dimensional tensor based on historical order data, independently embeds and encodes each type of travel mode by using an MLP (Multilayer Perceptron), and obtains low-dimensional input features. Subsequently, the low-dimensional input features are transmitted into a hybrid expert network composed of multiple "dedicated expert modules" and "dedicated expert modules", and a space-time joint feature extraction module composed of GCN (Graph Convolutional Network) and TCN (Temporal Convolutional Network) is embedded in the hybrid expert network. In order to adapt to the dynamics of the relationship between OD pairs, a graph structure learning mechanism is introduced into the GCN to adaptively optimize the adjacency matrix, thereby improving the generalization ability of the model to the spatial structure. In order to adapt to the differences in travel behavior at different times, the application designs a time-aware router, dynamically activates different sub-expert sets according to the space-time context, and weights and fuses the expert outputs through a gating module controlled by the context provided by the dedicated expert module. The fused features are sent to an enhanced quantile prediction module, which constructs multiple upper quantiles based on the median by using non-negative residual recursion, ensures the monotonicity of the prediction interval, and effectively avoids the quantile crossing problem.

[0041] In the design of the loss function, the application introduces a Huber smooth quantile loss function to improve the training stability, and combines a sample reweighting mechanism to suppress the zero sample dominance phenomenon and improve the generalization ability of the model in sparse scenarios. Finally, the model output is a set of quantile prediction results with different confidence levels, which supports the platform to realize more forward-looking intelligent scheduling and risk perception. The application has carried out empirical evaluation on multiple real data sets, and the results show that it is significantly better than existing mainstream models in terms of average error, coverage rate and stability, and has good practicality and popularization value.

[0042] The processing flow of the multi-mode online car-hailing demand uncertainty prediction method based on the multi-expert hybrid network provided by the embodiment of the application is as shown in Figure 1 The processing flow of the multi-mode online car-hailing demand uncertainty prediction method based on the multi-expert hybrid network provided by the embodiment of the application is as shown in

[0043] Step S10: Construct a hybrid expert network structure composed of dedicated expert modules and shared expert modules, and each dedicated expert module and shared expert module includes a graph convolutional network (GCN), a graph structure learning module (GSL) and a temporal convolutional network (TCN).

[0044] Step S20: Obtain historical OD demand data of a city multi-mode online car-hailing platform, and construct an OD three-dimensional tensor including multiple travel modes (such as express cars and carpooling).

[0045] Step S30: The OD three-dimensional tensor of multiple travel modes is taken as training data, and the hybrid expert network structure is trained by using the training data to obtain a trained hybrid expert network structure.

[0046] Step S40: The input data of the trained hybrid expert network structure is an OD three-dimensional tensor representing multiple travel models in a historical period, which can be mathematically calculated as Wherein T in represents the length of the time slice of the input hybrid expert network, N represents the number of OD pairs, and M represents the number of travel mode categories.

[0047] The output data of the trained hybrid expert network structure is an OD three-dimensional tensor representing multiple travel models in a future multiple time slices, which can be mathematically calculated as a set Wherein Wherein T out represents the length of the time slice of the output result of the hybrid expert network, N represents the number of OD pairs, and M represents the number of travel mode categories; τ is a quantile point, and T is a predefined quantile point set.

[0048] Because the neural network generally allows processing multiple samples when processing data, that is, input in batches. Therefore, the input data generally increases a batch size dimension. Therefore, the batch-packaged input and output can be calculated as four-dimensional tensors and four-dimensional tensor sets And Wherein

[0049] Specifically, the above step S10 comprises:

[0050] The hybrid expert network structure realizes the precision prediction and uncertainty quantification of urban multi-mode OD travel demand through joint learning, Figure 2 A module structure diagram of a time-aware hybrid expert network structure provided by the present application. The time-aware hybrid expert network is composed of exclusive expert modules, shared expert modules, gate modules and quantile point prediction modules,

[0051] Each exclusive expert module and shared expert module internally contains a space-time feature extraction submodule (ST-Block) composed of GCN and TCN, and a GSL (Graph Structure Learning) mechanism is embedded in the GCN, which is used to reconstruct an adaptive adjacency matrix through singular value decomposition to enhance the spatial modeling capability. The TCN part uses a multi-scale sliding window input to extract historical periodicity and short-term trends.

[0052] Shared expert module, used to accept joint input of all travel modes, learn inter-modal shared knowledge; the input of the shared expert module is data of all modes. During internal processing of the neural network, such data can be homogenized into a three-dimensional tensor TCN performs multi-layer time-domain convolution operation on T in dimension to obtain time hidden encoding H t ; GCN performs multi-layer graph convolution operation on N dimension combined with graph structure learning technology to obtain spatial hidden encoding H s . H t and H s are processed using multi-layer perceptron on T in dimension and N dimension respectively, and then the Hadamard product of the processing results is taken and processed again using multi-layer perceptron to obtain the spatio-temporal encoding output of the shared expert.

[0053] Exclusive expert module, used to receive input of only a specific travel mode, and model the specific travel mode. The input of the exclusive expert module is data of a single mode. During internal processing of the neural network, the data of the specific mode is first extracted, that is, an indexing operation is performed on M dimension to obtain a homogenized batch tensor TCN performs multi-layer time-domain convolution operation on T in dimension to obtain time hidden encoding H t ; GCN performs multi-layer graph convolution operation on N dimension combined with graph structure learning technology to obtain spatial hidden encoding H s . H t and H s are processed using multi-layer perceptron on T in dimension and N dimension respectively, and then the Hadamard product of the processing results is taken and processed again using multi-layer perceptron to obtain the spatio-temporal encoding output of the exclusive expert.

[0054] Gating module, used to dynamically activate different expert module combinations according to the current spatio-temporal context (such as morning and evening peak) by using a time-aware router; the output data from different expert modules are weighted and fused, and the weighted and fused data are transmitted to an enhanced quantile prediction module. The gating module is controlled by context information provided by the exclusive expert module. The gating module introduces a time-aware router module, dynamically determines the currently activated expert module combination according to the input spatio-temporal context information (such as peak / flat, weekday / holiday, etc.), thereby supporting dynamic modeling of time-varying tasks. This mechanism can be regarded as a soft gating selection mechanism, and has context adaptive capability.

[0055] The enhanced quantile prediction module is used for generating non-crossing multi-quantile prediction values satisfying monotonicity based on the weighted fused output data and the median regression result by using a residual construction mechanism, and deriving prediction interval results under each confidence level. The enhanced quantile prediction module uses a Huber smooth quantile loss function and introduces a zero sample re-weighting mechanism for training, so as to improve the prediction stability and accuracy of the model under the conditions of data sparsity and uncertainty.

[0056] Specifically, the step S20 includes:

[0057] The travel modes can be a fast car, a shared car and a taxi, etc. The historical OD sequences of each travel mode are obtained, the historical OD sequences of each travel mode are normalized and independently embedded by using an MLP, the hidden layer representation of the MLP is obtained, the OD three-dimensional tensor of each travel mode is constructed, and the low-dimensional input feature is obtained.

[0058] Specifically, the step S30 includes:

[0059] The OD three-dimensional tensors of multiple travel modes are taken as training data, and the training data is input into the hybrid expert network structure. The structure of a sliding window of a TCN in a space-time feature extraction submodule provided by the embodiment of the application is as shown in Figure 3 The sliding window of the TCN includes a neighboring time slice, a day-period time slice and a week-period time slice.

[0060] The gating module activates different expert module sets according to the space-time context information from the time-aware router, wherein the space-time context includes but is not limited to scenario labels such as weekdays and weekends, morning and evening peak and flat peak.

[0061] The implementation principle diagram of an enhanced quantile prediction module provided by the embodiment of the application is as shown in Figure 4 The enhanced quantile prediction module is based on the median regression value, accumulates non-negative offsets (generated by a ReLU activation function) in a recursive form, constructs multiple upper and lower quantile prediction values, and explicitly satisfies the monotonicity constraint between quantiles. The enhanced quantile prediction module is composed of a group of parallel MLP+ReLU structures, each layer prediction residual term δ k Each quantile value is generated by weighting recursion from the median prediction result

[0062] ​The enhanced quantile prediction module ensures that the quantile prediction meets the mathematically required monotonicity property (the upper quantile point is definitely not smaller than the lower quantile point), and avoids quantile crossing without post-processing. In addition, a residual connection mechanism is introduced, so that each layer of quantile prediction can reuse the information of the previous layer, reduce error accumulation and improve the stability of gradient propagation, thereby improving the numerical convergence performance in the training process. Each mode of travel (such as express car and carpooling) is equipped with an independent enhanced quantile prediction module to ensure that its prediction distribution reflects the behavior characteristics specific to the service, while improving the overall prediction stability in joint learning.

[0063] The final output of the above hybrid expert network structure is the OD prediction demand value of each travel mode under multiple confidence intervals, including the median prediction value and the prediction composed of quantiles.

[0064] OD three-dimensional tensor data of travel mode After inputting into the above hybrid expert network structure, first, an initial hidden layer encoding is obtained through a simple linear layer mapping. Secondly, two branches are passed through, which are the exclusive expert branch and the shared expert module. The exclusive expert branch extracts and calculates the spatio-temporal features of the hidden layer encoding, and finally gives the exclusive expert spatio-temporal encoding, which is simply denoted as exclusive encoding. The exclusive encoding is then input into the gating module. The shared expert module first determines the active subset of shared experts according to the time type of the data (such as whether it is a weekday, whether it is an early peak / late peak, etc.). Secondly, the active subset of shared experts will respectively extract and calculate the spatio-temporal features of the hidden layer encoding, and finally give their respective shared spatio-temporal encoding. The final shared spatio-temporal encoding is input into the gating module after being packed.

[0065] The gating module will refer to the spatio-temporal context of the exclusive encoding, adaptively adjust the weight of different shared spatio-temporal encodings, and finally fuse them with the exclusive encoding to obtain the final spatio-temporal expert encoding, which is denoted as fused spatio-temporal encoding.

[0066] The fused spatio-temporal encoding is finally input into the enhanced quantile prediction module, and a set of four-dimensional tensors

[0067] The training loss function of the above hybrid expert network structure is the Huber smoothed quantile loss, which adopts a sample re-weighting strategy for zero inflation data. Different gradient scaling factors are set for zero samples and non-zero samples to avoid model bias towards invalid areas. The definition is as follows:

[0068] (1) The quantile loss is defined as

[0069] (2) The smoothed Huber loss function is defined as

[0070] (3) Introduce a weighting mechanism for zero inflation characteristics, define the final loss as:

[0071] Wherein

[0072] The sample count operator represents, and is a minimum positive constant (such as 0.001), used to prevent division by zero. The above loss function form improves the numerical stability in the training process, effectively alleviates the prediction of quantile cross problem and zero sample dominant problem.

[0073] The value of the above training loss function is used for back propagation, and the parameters of the above hybrid expert network structure are iteratively updated. Until the change range of the value of the training loss function is less than the set threshold range, or the predetermined training termination condition is reached, the training of the above hybrid expert network structure is terminated, and the trained hybrid expert network structure is obtained.

[0074] The present application is essentially a computer program that needs to be deployed and run on a computer, implemented using the PyTorch software framework.

[0075] In summary, the present application considers the potential dynamic dependence relationship between different modes of the online car-hailing platform, uses a multi-expert hybrid neural network framework to model the demand transfer between different modes; uses graph structure learning, expert routing and other technologies to enhance the prediction performance and generalization ability of the model under different spatio-temporal context scenarios.

[0076] Those skilled in the art can understand that the drawings are only schematic diagrams of an embodiment, and the modules or processes in the drawings are not necessarily required to implement the present application.

[0077] From the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software and the necessary general hardware platform. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including a number of instructions to make a computer device (which can be a personal computer, server, or network device, etc.) execute the methods described in various embodiments or some parts of the embodiments.

[0078] The various embodiments described in this specification are described with reference to a particular sequence or order, but the order of the various steps comprising the embodiments can be modified so that particular sequences or orders make no significant difference in the final result or outcome of the methods taught herein. Additionally, some steps can be performed concurrently in parallel, rather than sequentially as described. The various embodiments can also be implemented by hardware components, software components, and / or combinations of hardware and software components.

[0079] The above descriptions are only the preferred embodiments of the present application, but the protection scope of the present application is not limited to this, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for multi-mode ride-hailing demand uncertainty prediction, characterized in that, include: A hybrid expert network structure consisting of dedicated expert modules and shared expert modules is constructed. Each dedicated expert module and shared expert module includes: Graph Convolutional Network (GCN), Graph Structure Learning Module (GSL), and Temporal Convolutional Network (TCN). Obtain historical origin-to-destination demand data from urban multi-modal ride-hailing platforms and construct an OD three-dimensional tensor that includes multiple travel modes; The OD three-dimensional tensors of multiple travel modes are used as training data, and the hybrid expert network structure is trained using the training data to obtain the trained hybrid expert network structure. The OD 3D tensors of multiple travel models over a historical period are input into the trained hybrid expert network structure. The trained hybrid expert network structure outputs the predicted OD 3D tensors of multiple travel modes in future time slices. The construction of the hybrid expert network structure consists of dedicated expert modules and shared expert modules. Each dedicated expert module and shared expert module includes: a graph convolutional network (GCN), a graph structure learning module (GSL), and a temporal convolutional network (TCN), comprising: A time-aware hybrid expert network is constructed, consisting of a dedicated expert module, a shared expert module, a gating module, and a quantile prediction module. Each dedicated expert module and shared expert module contains a space-time feature extraction submodule jointly composed of GCN and TCN. A graph structure learning GSL mechanism is embedded in the GCN, and an adaptive adjacency matrix is ​​reconstructed through singular value decomposition. The shared expert module is used to accept joint input from all travel modes and learn knowledge shared between modes; the dedicated expert module is used to receive input only for a specific travel mode and model that specific travel mode. The gating module is used to dynamically activate different expert module combinations based on the current spatiotemporal context using a time-aware router; to perform weighted fusion on the output data from different expert modules; and to transmit the weighted fused data to the enhanced quantile prediction module. The enhanced quantile prediction module is used to generate non-crossing multiquantile prediction values ​​that satisfy monotonicity based on the weighted fusion output data and the median regression results using the residual construction mechanism, and to derive the prediction interval results at each confidence level. The process of acquiring historical origin-to-destination (OD) demand data from urban multi-modal ride-hailing platforms and constructing an OD three-dimensional tensor including multiple travel modes includes: The historical OD sequence of each travel mode is obtained, the historical OD sequence of each travel mode is normalized and independently embedded by using a multi-layer perception MLP, a hidden layer representation of the MLP is obtained, and an OD three-dimensional tensor of each travel mode is constructed. The expression of the OD three-dimensional tensor is as follows: wherein represents the length of a historical time slice input into the mixed expert network, represents the number of OD pairs, represents the number of travel mode categories, and the travel modes include express cars, shared cars and taxis.

2. The method of claim 1, wherein, The method of using the OD three-dimensional tensors of multiple travel modes as training data, and using the training data to train the hybrid expert network structure to obtain the trained hybrid expert network structure, includes: OD three-dimensional tensor of multiple travel modes As the training data, after inputting the training data into the mixed expert network structure, firstly, an initial hidden layer code is obtained through a linear layer mapping, the hidden layer code is transmitted to the exclusive expert branch and the shared expert module, the exclusive expert branch performs space-time feature extraction and calculation on the hidden layer code, and outputs the exclusive expert space-time code, the exclusive expert space-time code is then input into the gating module, the shared expert module firstly determines the active shared expert subset according to the time type to which the data belongs, the active shared expert subset will respectively perform space-time feature extraction and calculation on the hidden layer code, and finally give the respective shared space-time code, and the final shared space-time code is input into the gating module after being packed. The gating module adaptively adjusts the weight of different shared spatiotemporal codes according to the spatiotemporal context of the dedicated code, and finally merges the shared spatiotemporal codes with the dedicated expert spatiotemporal codes to obtain the fused spatiotemporal code. The spatiotemporal context includes, but is not limited to, weekday and weekend, morning and evening peak and off-peak scene labels. inputting the fused spatio-temporal encoding into an enhanced quantile prediction module, the enhanced quantile prediction module being composed of a set of parallel MLP + ReLU structures, each layer predicting a residual term , each quantile value being generated by recursively weighting the median prediction result , the enhanced quantile prediction module being based on the median regression value, constructing a plurality of upper and lower quantile prediction values by recursively accumulating a non-negative offset, and outputting the OD prediction demand value of each travel mode under a plurality of confidence intervals;​ The training loss function of the mixed expert network structure is Huber smooth quantile loss, the value of the training loss function is used for back propagation, and the parameters of the mixed expert network structure are iteratively updated until the change range of the value of the training loss function is less than a set threshold range or a predetermined training termination condition is reached, the training of the mixed expert network structure is terminated, and a trained mixed expert network structure is obtained.

3. The method of claim 2, wherein, The method further comprises: a shared expert module for accepting the OD three-dimensional tensor of all travel modes , learning inter-modal shared knowledge, the TCN in the shared expert module performs multi-layer time domain convolution operation in the dimension to obtain time hidden encoding , the GCN in the shared expert module combines graph structure learning technology to perform multi-layer graph convolution operation in the dimension to obtain spatial hidden encoding , and respectively use multi-layer perceptron for processing in the dimension and the dimension, then take the Hadamard product of the processing results and use multi-layer perceptron again for processing to obtain shared spatio-temporal encoding of the shared expert; A dedicated expert module for receiving only OD tensors of a specific travel mode Modeling specific travel modes, the TCN in the dedicated expert module performs multi-layer temporal convolution operations on the dimension to obtain temporal hidden encodings The GCN in the dedicated expert module performs multi-layer graph convolution operations on the dimension in combination with graph structure learning techniques to obtain spatial hidden encodings , and are processed using multi-layer perceptrons on the dimension and the dimension, respectively, and then the Hadamard product of the processing results is taken and processed again using a multi-layer perceptron to obtain the dedicated expert spatiotemporal encoding.

4. The method according to claim 2 or 3, characterized in that, The method further comprises: The method further comprises: The method further comprises: obtaining an OD three-dimensional tensor of a plurality of trip models in a historical time , representing a length of a time slice in the historical time, representing a number of OD pairs, representing a number of trip pattern categories, inputting the OD three-dimensional tensor of the plurality of trip models into a trained hybrid expert network structure, an output of the trained hybrid expert network structure predicting an OD three-dimensional tensor set of a plurality of trip models in a future plurality of time slices wherein , representing a length of a time slice of the hybrid expert network output result, representing a number of OD pairs, representing a number of trip pattern categories, is a quantile point, is a predefined set of quantile points.

Images