Multi-mode online car-hailing demand uncertainty prediction method

Abstract

The invention provides a multi-mode online car-hailing demand uncertainty prediction method. The method comprises the following steps: constructing a hybrid expert network structure formed by an exclusive expert module and a shared expert module, obtaining OD demand data from a historical starting point to a terminal point of the urban multi-mode online car-hailing platform, and constructing an OD three-dimensional tensor comprising a plurality of travel modes; taking the OD three-dimensional tensors of the plurality of travel modes as training data, and training the hybrid expert network structure by using the training data to obtain a trained hybrid expert network structure; and inputting the OD three-dimensional tensors of the plurality of travel models in a period of historical time into the trained hybrid expert network structure, and predicting the OD three-dimensional tensors of the plurality of travel models in a future time slice. According to the method, through learning of the space-time dependency relationship and the operation dependency relationship, the accuracy of carpooling demand prediction is improved, and important support is provided for a platform to deploy vehicle resources more reasonably, meet user travel demands and improve service quality.

Description

Technical Field

[0001] The present invention relates to the technical field of transportation demand, and in particular to a method for predicting uncertainty in multi-modal online ride-hailing demand. Background Art

[0002] In recent years, with the development of the internet and mobile devices, the vast amount of order data recorded by online ride-hailing platforms has provided a rich source of training data for prediction models. This massive amount of data has enabled the widespread application of deep learning methods in the field of ride-sharing demand forecasting. Regarding data sources, researchers have begun to focus on integrating multi-source data and multi-dimensional features. By incorporating various external factors such as weather, architectural environment, and points of interest, they have constructed more complex spatiotemporal prediction models.

[0003] One existing method for forecasting ride-hailing demand includes the STZINB-GNN model. This model is a probabilistic forecasting model that incorporates the zero-inflated negative binomial (ZINB) distribution and enables forecasting of ride-hailing demand in a city. This model incorporates a zero-inflated statistical model and has the ability to quantify uncertainty.

[0004] The shortcomings of one of the above-mentioned methods for predicting online ride-hailing demand in the prior art include: the online ride-hailing demand prediction models in the prior art mostly focus on a single service mode (such as a private car) and cannot capture the interactive relationship between multiple travel modes in the city. Especially in urban ride-sharing (carpooling) services, there are obvious sparsity, zero expansion and time heterogeneity, which make it difficult for existing deep learning models to accurately model and predict. At the same time, most existing methods focus on the "point prediction" of the average demand value, lack effective quantification of prediction uncertainty, and are difficult to support the platform's robust scheduling and risk perception needs. Summary of the Invention

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

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

[0007] A method for predicting uncertainty in multi-modal ride-hailing demand, comprising:

[0008] Construct a hybrid expert network structure consisting of dedicated expert modules and shared expert modules. Each dedicated expert module and shared expert module includes: graph convolutional network (GCN), graph structure learning module (GSL), and temporal convolutional network (TCN).

[0009] Obtain historical origin-to-destination OD demand data from a city's multi-modal ride-hailing platform and construct a three-dimensional OD tensor encompassing multiple travel modes.

[0010] 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 a trained hybrid expert network structure;

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

[0012] Preferably, the hybrid expert network structure is constructed by a dedicated expert module and a shared expert module, each of which includes: a graph convolutional network GCN, a graph structure learning module GSL and a temporal convolutional network TCN, including:

[0013] 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 spatial-temporal feature extraction submodule composed of a GCN and a TCN. The GCN is embedded with a graph structure learning (GSL) mechanism, and the adaptive adjacency matrix is ​​reconstructed through singular value decomposition.

[0014] The shared expert module is used to accept the joint input of all travel modes and learn the shared knowledge between modes; the exclusive expert module is used to only receive the input of a specific travel mode and model the specific travel mode.

[0015] The gating module is used to dynamically activate different expert module combinations according to the current spatiotemporal context using a time-aware router; perform weighted fusion on the output data from different expert modules, and transmit the weighted fusion data to the enhanced quantile prediction module;

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

[0017] Preferably, the acquisition of historical start-to-end point OD demand data of a city's multi-modal online ride-hailing platform and the construction of a three-dimensional OD tensor including multiple travel modes include:

[0018] Obtain the historical OD sequence of each travel mode, use the multi-layer perceptron (MLP) to normalize and independently embed the historical OD sequence of each travel mode, obtain the hidden layer representation of the MLP, and construct the OD three-dimensional tensor of each travel mode. The expression of the OD three-dimensional tensor is: Where T inrepresents the length of the historical time slice input into the hybrid expert network, N represents the number of OD pairs, and m represents the number of travel mode types, including express, carpooling, and taxi.

[0019] Preferably, 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 a trained hybrid expert network structure, including:

[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 first passes through a linear layer mapping to obtain an initial hidden layer code, which is then transmitted to the exclusive expert branch and the shared expert module. The exclusive expert branch extracts and calculates the spatiotemporal features of the hidden layer code and outputs the exclusive expert spatiotemporal code. The exclusive expert spatiotemporal code is then input into the gating module. The shared expert module first determines the active shared expert subset according to the time type of the data. The active shared expert subsets will respectively extract and calculate the spatiotemporal features of the hidden layer code, and finally give their own shared spatiotemporal codes. The final shared spatiotemporal codes are packaged and input into the gating module.

[0021] The gating module adaptively adjusts the weights of different shared spatiotemporal codes according to the spatiotemporal context of the dedicated code, and finally fuses the shared spatiotemporal code with the dedicated expert spatiotemporal code to obtain a fused spatiotemporal code. The spatiotemporal context includes but is not limited to weekday and weekend, morning and evening peak and off-peak scenario labels;

[0022] The fused spatiotemporal coding is input into the enhanced quantile prediction module, which consists of a set of parallel MLP+ReLU structures. Each layer predicts the residual term δ k , each quantile value is predicted by the median Weighted recursive generation: The enhanced quantile prediction module constructs multiple upper and lower quantile prediction values ​​based on the median regression value by recursively accumulating non-negative offsets, 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 the Huber smoothed quantile loss. The value of the training loss function is used for back propagation, and the parameters of the training of the above-mentioned hybrid expert network structure are iteratively updated until the variation 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 hybrid expert network structure is terminated, and a trained hybrid expert network structure is obtained.

[0024] Preferably, the method further comprises:

[0025] Shared expert module that accepts OD 3D tensors for all travel modes Learning models share knowledge, and TCN in the shared expert module is in Multi-layer time domain convolution operation is performed on the dimension to obtain the time hidden code H t , the GCN in the shared expert module combines the graph structure learning technology to perform multi-layer graph convolution operations in N dimensions to obtain the spatial hidden code H s , H t and H s Respectively in T in The dimension and N dimension are processed using a multi-layer perceptron, and then the Hadamard product of the processing results is processed again using a multi-layer perceptron to obtain the shared spatiotemporal encoding of the shared expert;

[0026] Dedicated expert module for receiving only the OD 3D tensor of a specific travel mode To model a specific travel mode, the TCN in the dedicated expert module is used in T in Multi-layer time domain convolution operation is performed on the dimension to obtain the time hidden code H t , the GCN in the dedicated expert module combines the graph structure learning technology to perform multi-layer graph convolution operations in N dimensions to obtain the spatial hidden code H s , H t and H s Respectively in T in The dimension and N dimension are processed using a multi-layer perceptron, and then the Hadamard product of the processing results is processed again using a multi-layer perceptron to obtain the exclusive expert spatiotemporal coding.

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

[0028] Get the OD three-dimensional tensor of multiple travel models in a historical time T in Represents the length of a time slice in a historical period, N represents the number of OD pairs, and M represents the number of travel mode types. The OD three-dimensional tensors X of the multiple travel models are input into the trained hybrid expert network structure. The output of the trained hybrid expert network structure predicts one of the multiple future time slices, which represents the OD three-dimensional tensor set of multiple travel models. in T outRepresents the time slice length of the output result of the hybrid expert network, N represents the number of OD pairs, M represents the number of travel mode types, τ is the quantile, and T is a predefined quantile set.

[0029] It can be seen from the technical solutions provided by the above-mentioned embodiments of the present invention that the present invention aims to predict the short-term demand for shared travel on a large scale of OD (Origin-Destination) by learning about spatiotemporal dependencies and operational dependencies, thereby improving the accuracy of shared travel demand predictions and providing important support for the platform to more reasonably allocate vehicle resources, meet user travel needs, and improve service quality.

[0030] Additional aspects and advantages of the present invention will be set forth in part in the following description, will become apparent from the following description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0032] Figure 1 A processing flow chart of a method for predicting uncertainty in multi-modal online ride-hailing demand provided by an embodiment of the present invention;

[0033] Figure 2 A module structure diagram of a time-aware hybrid expert network structure provided by an embodiment of the present invention;

[0034] Figure 3 A structural diagram of a TCN sliding window in a space-time feature extraction submodule provided in an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the implementation principle of an enhanced quantile prediction module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0037] It will be understood by those skilled in the art that, unless expressly stated otherwise, the singular forms "a", "an", "said" and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the description of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or couplings. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0038] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which the present invention pertains. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as such herein, will not be interpreted in an idealized or overly formal sense.

[0039] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0040] The method of the present invention first constructs a multi-modal OD three-dimensional tensor based on historical order data, and uses MLP (Multilayer Perceptron) to independently embed and encode each type of travel mode to obtain low-dimensional input features. Subsequently, the low-dimensional input features are passed into a hybrid expert network composed of multiple "exclusive expert modules" and "exclusive expert modules". The hybrid expert network embeds a space-time joint feature extraction module composed of GCN (Graph Convolutional Network) and TCN (Temporal Convolutional Network). In order to adapt to the dynamic nature of the relationship between OD pairs, a graph structure learning mechanism is introduced into GCN to adaptively optimize the adjacency matrix, thereby improving the model's generalization ability for spatial structures. In order to adapt to the differences in travel behavior in different time periods, the present invention designs a time-aware router, dynamically activates different sub-expert sets according to the spatiotemporal context, and performs weighted fusion of expert outputs through a gating module with context control provided by the exclusive expert module. The fused features are fed into the enhanced quantile prediction module, which takes the median as the benchmark and recursively constructs multiple upper and lower quantiles through non-negative residuals to ensure the monotonicity of the prediction interval and effectively avoid the quantile crossing problem.

[0041] In terms of loss function design, this paper introduces the Huber smoothed quantile loss function to improve training stability, while combining it with a sample reweighting mechanism to suppress the zero-sample dominance phenomenon and improve the model's generalization ability in sparse scenarios. Ultimately, the model outputs a set of quantile prediction results with different confidence levels, supporting the platform to achieve more forward-looking intelligent scheduling and risk perception. This paper has been empirically evaluated on multiple real-world datasets, and the results show that it significantly outperforms existing mainstream models in terms of average error, coverage, and stability, demonstrating good practicality and promotion value.

[0042] The processing flow of a multi-mode online car-hailing demand uncertainty prediction method based on a multi-expert hybrid network provided by an embodiment of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0043] Step S10: Construct a hybrid expert network structure consisting of a dedicated expert module and a shared expert module. 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 the city's multi-modal online car-hailing platform and construct an OD three-dimensional tensor including multiple travel modes (such as express and carpooling).

[0045] Step S30: 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 a trained hybrid expert network structure.

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

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

[0048] Because neural networks generally allow processing of multiple samples when processing data, that is, input in batches. Therefore, the input data generally increases a batch size dimension. Therefore, the input and output after batch packaging can be counted as a four-dimensional tensor and a four-dimensional tensor set respectively. and in

[0049] Specifically, the above step S10 includes:

[0050] The hybrid expert network structure achieves accurate prediction and uncertainty quantification of urban multi-modal OD travel demand through joint learning. Figure 2 This is a module structure diagram of a time-aware hybrid expert network structure provided by the present invention. The time-aware hybrid expert network consists of a dedicated expert module, a shared expert module, a gating module, and a quantile prediction module.

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

[0052] The shared expert module is used to accept the joint input of all travel modes and learn the shared knowledge between modes; the input of the shared expert module is the data of all modes. In the internal processing of the neural network, this kind of data can be isomorphic to a three-dimensional tensor TCN in T in Multi-layer time domain convolution operation is performed on the dimension to obtain the time hidden code H t ; GCN combines graph structure learning technology in N dimensions to perform multi-layer graph convolution operations to obtain spatial hidden coding H s . H t and H s Respectively in T in The dimension and N dimension are processed using a multi-layer perceptron, and then the Hadamard product of the processing results is processed again using a multi-layer perceptron to obtain the spatiotemporal encoding output of the shared expert.

[0053] The dedicated expert module is used to receive input from a specific travel mode and model the specific travel mode. The input of the dedicated expert module is the data of a single mode. During the internal processing of the neural network, the data of the specific mode is first extracted, that is, the index operation is performed on the M dimension to obtain a homogeneous batch tensor. TCN in T in Multi-layer time domain convolution operation is performed on the dimension to obtain the time hidden code H t ; GCN combines graph structure learning technology in N dimensions to perform multi-layer graph convolution operations to obtain spatial hidden coding H s . H t and H s Respectively in T in The dimension and N dimension are processed using a multi-layer perceptron, and then the Hadamard product of the processing results is processed again using a multi-layer perceptron to obtain the spatiotemporal coding output of the dedicated expert.

[0054] The gating module utilizes a time-aware router to dynamically activate different combinations of expert modules based on the current spatiotemporal context (e.g., morning and evening rush hour). The output data from different expert modules is weighted and fused, and the weighted fusion data is transmitted to the enhanced quantile prediction module. The gating module is controlled by contextual information provided by dedicated expert modules. The gating module incorporates a time-aware router module, which dynamically determines the currently activated expert module combination based on the input spatiotemporal context (e.g., peak / off-peak, weekday / holiday), thereby supporting dynamic modeling of time-varying tasks. This mechanism can be considered a soft gating selection mechanism with contextual adaptability.

[0055] The enhanced quantile prediction module utilizes a residual construction mechanism to generate monotonic, non-intersecting multi-quantile predictions based on weighted fusion output data and median regression results, and derives prediction intervals at various confidence levels. This module uses the Huber smoothed quantile loss function and introduces a zero-sample reweighting mechanism for training to improve the model's prediction stability and accuracy under conditions of data sparsity and uncertainty.

[0056] Specifically, the above step S20 includes:

[0057] The above travel modes can be express, carpooling, and taxis. The historical OD sequence of each travel mode is obtained and normalized and independently embedded using an MLP to obtain the hidden layer representation of the MLP. A three-dimensional OD tensor of each travel mode is constructed to obtain low-dimensional input features.

[0058] Specifically, the above step S30 includes:

[0059] The OD three-dimensional tensors of multiple travel modes are used as training data, and the training data is input into the above hybrid expert network structure. The structure of the sliding window of TCN in the space-time feature extraction submodule provided by the embodiment of the present invention is as follows Figure 3 As shown in FIG, the sliding window of TCN includes: adjacent time slices, daily-cycle time slices and weekly-cycle time slices.

[0060] The gating module activates different sets of expert modules through the time-aware router according to the spatiotemporal context information, where the spatiotemporal context includes but is not limited to situational labels such as weekdays and weekends, morning and evening peaks and off-peak hours.

[0061] The implementation principle diagram of an enhanced quantile prediction module provided by an embodiment of the present invention is as follows: Figure 4 As shown in Figure 1, the enhanced quantile prediction module is based on the median regression value, and recursively accumulates non-negative offsets (generated by the ReLU activation function) to construct multiple upper and lower quantile prediction values, and explicitly satisfies the monotonicity constraint between quantiles. The enhanced quantile prediction module consists of a set of parallel MLP+ReLU structures, and each layer predicts the residual term δ k , each quantile value is predicted by the median Weighted recursive generation:

[0062] The enhanced quantile prediction module ensures that quantile predictions meet the mathematically required monotonicity property (the upper quantile must be no less than the lower quantile), avoiding quantile crossover without the need for post-processing. Furthermore, a residual connection mechanism is introduced, allowing each layer of quantile prediction to reuse information from the previous layer, reducing error accumulation and improving gradient propagation stability, thereby enhancing numerical convergence performance during training. Each travel mode (such as express and carpooling) is equipped with a separate enhanced quantile prediction module to ensure that its predicted distribution reflects the unique behavioral characteristics of the service, while also improving overall prediction stability in joint learning.

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

[0064] OD three-dimensional tensor data of travel mode After being input into the above-mentioned hybrid expert network structure, it first passes through a simple linear layer mapping to obtain the initial hidden layer code. Secondly, it passes through two branches, namely the exclusive expert branch and the shared expert module. The exclusive expert branch extracts and calculates the spatiotemporal features of the hidden layer code, and finally gives the exclusive expert spatiotemporal code, which is simply referred to as the exclusive code. The exclusive code is then input into the gating module. The shared expert module first determines the active shared expert subset based on the time type of the data (such as whether it is a weekday, whether it is morning rush hour / evening rush hour, etc.). Secondly, the active shared expert subsets will extract and calculate the spatiotemporal features of the hidden layer code respectively, and finally give their own shared spatiotemporal codes. The final shared spatiotemporal code is packaged and input into the gating module.

[0065] The gating module will refer to the spatiotemporal context of the dedicated code, adaptively adjust the weights of different shared spatiotemporal codes, and finally fuse them with the dedicated code to obtain the final spatiotemporal expert code, which is counted as the fused spatiotemporal code.

[0066] The fused spatiotemporal coding is finally input into the enhanced quantile prediction module, and the four-dimensional tensor set is obtained after module processing.

[0067] The training loss function of the hybrid expert network structure is Huber smoothed quantile loss. A sample reweighting strategy is used for zero-inflated data, where different gradient scaling factors are set for zero samples and non-zero samples to avoid the model from being biased into invalid areas. The definition is as follows:

[0068] (1) Quantile loss is defined as

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

[0070] (3) A weighting mechanism is introduced to address the zero-expansion feature, and the final loss is defined as:

[0071] in

[0072] represents the sample counting operator, and ∈ is a very small positive constant (such as 0.001) to prevent division by zero. The above loss function form improves numerical stability during training and effectively alleviates the problem of quantile crossing and zero sample dominance in prediction.

[0073] Backpropagation is performed using the value of the training loss function to iteratively update the parameters of the hybrid expert network structure. Training of the hybrid expert network structure is terminated until the value of the training loss function falls below a set threshold or a predetermined training termination condition is met, resulting in a trained hybrid expert network structure.

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

[0075] In summary, the present invention takes into account the potential dynamic dependencies between different modes of the online ride-hailing platform, and uses a multi-expert hybrid neural network framework to model the demand transfer between different modes; it adopts graph structure learning, expert routing and other technologies to enhance the model's predictive performance and generalization capabilities in different spatiotemporal context scenarios.

[0076] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0077] From the above description of the embodiments, it can be seen that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus the necessary general-purpose hardware platform. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0078] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiments. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without making any creative efforts.

[0079] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for predicting uncertainty in multi-modal ride-hailing demand, characterized by: include: Construct a hybrid expert network structure consisting of dedicated expert modules and shared expert modules. 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 OD demand data from a city's multi-modal ride-hailing platform and construct a three-dimensional OD tensor encompassing 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 a trained hybrid expert network structure; The OD three-dimensional tensors of multiple travel models in a historical time period are input into the trained hybrid expert network structure, and the trained hybrid expert network structure outputs the predicted OD three-dimensional tensors of multiple travel modes in a future time slice.

2. The method according to claim 1, characterized in that The hybrid expert network structure is constructed by a dedicated expert module and a shared expert module. 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, including: 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 spatial-temporal feature extraction submodule composed of a GCN and a TCN. The GCN is embedded with a graph structure learning (GSL) mechanism, and the adaptive adjacency matrix is ​​reconstructed through singular value decomposition. The shared expert module is used to accept the joint input of all travel modes and learn the shared knowledge between modes; the exclusive expert module is used to only receive the input of a specific travel mode and model the specific travel mode. The gating module is used to dynamically activate different expert module combinations according to the current spatiotemporal context using a time-aware router; perform weighted fusion on the output data from different expert modules, and transmit the weighted fusion data to the enhanced quantile prediction module; The enhanced quantile prediction module is used to generate non-crossing multi-quantile prediction values ​​that meet monotonicity based on the weighted fused output data and the median regression results using a residual construction mechanism, and derive prediction interval results at each confidence level.

3. The method according to claim 1, characterized in that The method of obtaining historical start-to-end point OD demand data of a city's multi-modal online ride-hailing platform and constructing a three-dimensional OD tensor including multiple travel modes includes: Obtain the historical OD sequence of each travel mode, use the multi-layer perceptron (MLP) to normalize and independently embed the historical OD sequence of each travel mode, obtain the hidden layer representation of the MLP, and construct the OD three-dimensional tensor of each travel mode. The expression of the OD three-dimensional tensor is: Where T in represents the length of the historical time slice input into the hybrid expert network, N represents the number of OD pairs, and M represents the number of travel mode types, including express, carpooling, and taxi.

4. The method according to claim 3, characterized in that 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 a trained hybrid expert network structure includes: 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 first passes through a linear layer mapping to obtain an initial hidden layer code, which is then transmitted to the exclusive expert branch and the shared expert module. The exclusive expert branch extracts and calculates the spatiotemporal features of the hidden layer code and outputs the exclusive expert spatiotemporal code. The exclusive expert spatiotemporal code is then input into the gating module. The shared expert module first determines the active shared expert subset according to the time type of the data. The active shared expert subsets will respectively extract and calculate the spatiotemporal features of the hidden layer code, and finally give their own shared spatiotemporal codes. The final shared spatiotemporal codes are packaged and input into the gating module. The gating module adaptively adjusts the weights of different shared spatiotemporal codes according to the spatiotemporal context of the dedicated code, and finally fuses the shared spatiotemporal code with the dedicated expert spatiotemporal code to obtain a fused spatiotemporal code. The spatiotemporal context includes but is not limited to weekday and weekend, morning and evening peak and off-peak scenario labels; The fused spatiotemporal coding is input into the enhanced quantile prediction module, which consists of a set of parallel MLP+ReLU structures. Each layer predicts the residual term δ k , each quantile value is predicted by the median Weighted recursive generation: The enhanced quantile prediction module constructs multiple upper and lower quantile prediction values ​​based on the median regression value by recursively accumulating non-negative offsets, and outputs the OD prediction demand value of each travel mode under multiple confidence intervals; The training loss function of the hybrid expert network structure is the Huber smoothed quantile loss. The value of the training loss function is used for back propagation, and the parameters of the training of the above-mentioned hybrid expert network structure are iteratively updated until the variation 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 hybrid expert network structure is terminated, and a trained hybrid expert network structure is obtained.

5. The method according to claim 4, characterized in that The method further comprises: Shared expert module that accepts OD 3D tensors for all travel modes Learning models share knowledge, and TCN in the shared expert module is in Multi-layer time domain convolution operation is performed on the dimension to obtain the time hidden code H t , the GCN in the shared expert module combines the graph structure learning technology to perform multi-layer graph convolution operations in N dimensions to obtain the spatial hidden code H s , H t and H s Respectively in T in The dimension and N dimension are processed using a multi-layer perceptron, and then the Hadamard product of the processing results is processed again using a multi-layer perceptron to obtain the shared spatiotemporal encoding of the shared expert; Dedicated expert module for receiving only the OD 3D tensor of a specific travel mode To model a specific travel mode, the TCN in the dedicated expert module is used in T in Multi-layer time domain convolution operation is performed on the dimension to obtain the time hidden code H t , the GCN in the dedicated expert module combines the graph structure learning technology to perform multi-layer graph convolution operations in N dimensions to obtain the spatial hidden code H s , H t and H s Respectively in T in The dimension and N dimension are processed using a multi-layer perceptron, and then the Hadamard product of the processing results is processed again using a multi-layer perceptron to obtain the exclusive expert spatiotemporal coding.

6. The method according to claim 4 or 5, characterized in that The OD three-dimensional tensors of multiple travel models in a historical time period are input into the trained hybrid expert network structure, and the trained hybrid expert network structure outputs the OD three-dimensional tensors of multiple travel modes in the predicted future time slice, including: Get the OD three-dimensional tensor of multiple travel models in a historical time T in Represents the length of a time slice in a historical period, N represents the number of OD pairs, and M represents the number of travel mode types. The OD three-dimensional tensors X of the multiple travel models are input into the trained hybrid expert network structure. The output of the trained hybrid expert network structure predicts one of the multiple future time slices, which represents the OD three-dimensional tensor set of multiple travel models. in T out Represents the time slice length of the output result of the hybrid expert network, N represents the number of OD pairs, M represents the number of travel mode types, τ is the quantile, and T is a predefined quantile set.

Images