A resident travel demand prediction method and system based on user portrait

Abstract

The application relates to a resident travel demand prediction method and system based on a user portrait, which comprises the following steps: in response to a demand prediction instruction and travel starting point information, target region information and a target group are determined, and prediction constraint conditions are obtained; a prediction coordinate matrix and a corresponding selection probability matrix are generated through a prediction model, a continuous probability distribution graph is generated through a spatial interpolation strategy, and is rendered into a probability distribution heat map; in response to an interactive selection operation, geometric boundary information of the target region is determined; first point of interest information is identified based on the geometric boundary information, and industry distribution data is obtained; a prediction conclusion report is generated through a semantic analysis model; in summary, the application generates a probability distribution heat map in response to a demand prediction instruction and supports interactive selection, finally generates a prediction conclusion report, thereby predicting travel demand on a geographical scale basis, realizing interactive exploration and analysis integration, and having the effects of improving the efficiency and practicality of decision-making.

Description

Technical Field

[0001] This application relates to the technical field of urban planning, and in particular to a method and system for predicting residents' travel demand based on user profiles. Background Technology

[0002] In business intelligence and urban planning practices, accurately predicting the spatial distribution of travel destinations for specific groups is a crucial step in optimizing business layout and transportation resource allocation.

[0003] Existing technical solutions generally employ historical statistical data or questionnaire surveys for macro-level aggregation analysis of residents' travel. However, this macro-aggregation approach clearly suffers from the following drawbacks: First, the output of existing methods is limited to coarse-grained data at the district or street level, making it difficult to characterize the probability distribution of travel destinations at a fine geographical scale. This results in severely insufficient spatial resolution of the prediction results. For example, planning managers struggle to obtain the distribution of demand hotspots at the point of interest level, thus failing to distinguish the differences in attractiveness among different commercial facilities within the same area. This leads to a lack of micro-level basis for site selection decisions and a high risk of resource misallocation. Second, existing technologies lack interactive exploration mechanisms, preventing planning managers from dynamically adjusting key parameters such as travel time and mode of transportation. Furthermore, they cannot intuitively understand the real-time impact of parameter changes on destination distribution, nor can they conduct focused analysis on arbitrarily defined areas on a map to achieve in-depth perception in conjunction with local business characteristics.

[0004] In summary, existing static aggregation analysis models make it difficult for planning managers to cope with complex and ever-changing real-world scenarios. For example, in the scenario of evaluating the planning of new commercial districts, it is impossible to quickly verify changes in passenger flow distribution under different traffic conditions. More importantly, in the existing static model, the data analysis process, spatial visualization, and semantic interpretation outputs mentioned above usually require planning managers to manually integrate them, resulting in the fragmentation of each process and a break in the flow from raw data to decision support. This severely restricts the practicality and response efficiency of real-time travel forecast analysis. Summary of the Invention

[0005] To address the aforementioned shortcomings, this application provides a method and system for predicting residents' travel demand based on user profiles.

[0006] The above-mentioned objective of this application is achieved through the following technical solution: A method for predicting residents' travel demand based on user profiles includes the following steps: In response to the demand forecasting command input by the user and the corresponding travel origin information, the target area information and target group are determined, and the forecasting constraints associated with the demand forecasting command are obtained. The forecasting constraints include travel time information, travel mode information and travel population profile tags. The prediction constraints are input into the pre-trained prediction model to generate the prediction coordinate matrix and its corresponding selection probability matrix. Based on the predicted coordinate matrix and the selection probability matrix, a continuous probability distribution map associated with the target area information is generated through a predefined spatial interpolation strategy, and the continuous probability distribution map is rendered as a probability distribution heatmap in an interactive scene. In response to the user's interactive selection operation of the target area in the probability distribution heatmap, determine the geometric boundary information of the target area; Based on geometric boundary information, identify all first points of interest within the target area and obtain business distribution data associated with each first point of interest; The selected probability matrix, the first point of interest information, and the associated business distribution data are input into a pre-trained semantic analysis model to generate a prediction conclusion report.

[0007] The second objective of this invention is achieved through the following technical solution: A resident travel demand prediction system based on user profiles includes: The constraint acquisition module is used to respond to the demand prediction command input by the user terminal and the corresponding travel origin information, determine the target area information and target group, and acquire the prediction constraints associated with the demand prediction command. The prediction constraints include travel time information, travel mode information and travel population profile tags. The matrix generation module is used to input the prediction constraints into the pre-trained prediction model and generate the prediction coordinate matrix and its corresponding selection probability matrix. The heatmap generation module is used to generate a continuous probability distribution map associated with target area information based on the predicted coordinate matrix and the selection probability matrix through a predefined spatial interpolation strategy, and render the continuous probability distribution map as a probability distribution heatmap in an interactive scene. The boundary determination module is used to determine the geometric boundary information of the target area in response to the user's interactive selection operation on the probability distribution heatmap. The data acquisition module is used to identify all first points of interest within the target area based on geometric boundary information, and to acquire business distribution data associated with each first point of interest. The report generation module is used to input the selection probability matrix, the first point of interest information and its associated business distribution data into a pre-trained semantic analysis model to generate a prediction conclusion report.

[0008] In summary, the method and system for predicting resident travel demand based on user profiles provided in this application generate a probability distribution heatmap in response to demand prediction instructions and support interactive selection. Finally, it integrates and generates a prediction conclusion report, thereby predicting travel demand on a fine geographical scale, realizing interactive exploration and automatic analysis integration, and improving the efficiency and practicality of decision-making. Attached Figure Description

[0009] Figure 1 This is a flowchart of an embodiment of a method for predicting residents' travel demand based on user profiles according to this application; Figure 2 This is a flowchart of step S20 in an embodiment of a method for predicting residents' travel demand based on user profiles in this application. Detailed Implementation

[0010] The following is in conjunction with the appendix Figures 1-2 This application will be described in further detail.

[0011] This application discloses a method for predicting residents' travel demand based on user profiles. In one embodiment, as follows: Figure 1 As shown, the specific steps include the following: S10: In response to the demand prediction command input by the user terminal and the corresponding travel origin information, determine the target area information and target group, and obtain the prediction constraints associated with the demand prediction command. The prediction constraints include travel time information, travel mode information and travel population profile tags. In this embodiment, the demand forecasting instruction refers to the operation command or trigger signal initiated by the user terminal to request travel demand forecasting, which usually contains the basic intent or goal of the analysis; the travel origin information refers to the starting location of the travel activity to be analyzed, which is usually determined by the user terminal selecting the corresponding location on the map, and is the benchmark point for spatial forecasting; the target area information refers to the geographical range to be analyzed, determined based on the travel origin and other information or defined by the user, and the target area information is the spatial boundary for subsequent heat map rendering and insight analysis; the target group refers to the specific group of people targeted by this forecasting analysis; the forecasting constraints refer to the various parameters used to limit the forecasting range and conditions when forecasting travel demand, including travel time information, travel mode information, and traveler profile tags; the traveler profile tags refer to the classification labels used to classify, identify, and define specific user groups with similar social attributes, behavioral characteristics, or consumption preferences.

[0012] Specifically, in response to the user's input demand prediction command and corresponding travel origin information, the system determines the target area information and target group, and obtains the prediction constraints associated with the demand prediction command. These constraints include travel time information, travel mode information, and traveler profile tags. Specifically, users can input their needs through a pre-set interface, such as specifying the travel date, time period, departure location, and desired travel mode. Upon receiving the user's input, the system determines the target area information based on a preset geographical range or a range specified by the user, and filters out the target group that meets the criteria from a preset user database based on the user's input profile tags. Subsequently, the input and determined information are integrated into prediction constraints. For example, if a user inputs a demand such as "Tomorrow morning at 9:00, departing from the company, walking, target group is young white-collar workers," then the system determines the target area within a 5-kilometer radius of the company location, and filters out the target group from the user database based on the "young white-collar workers" tag, ultimately forming prediction constraints that include travel time, travel mode, and traveler profile tags.

[0013] It should be noted that the traveler profile labels in this embodiment are not simple information classifications, but feature vectors used for model calculation after technical processing. The specific processing steps can be exemplarily included as follows: User profile tag acquisition: User profile tags can be obtained from authorized data in mobile applications, including age group, occupation type, consumption level, travel preferences, etc.

[0014] Encoding of profile tags: For discrete tags, such as age group and occupation type, a learnable embedding matrix is ​​used to map them into a 128-dimensional dense vector. This embedding matrix can be automatically learned through the neural network training process to capture the potential semantic associations between different tags. For example, the association strength between students and libraries is higher than that between students and bars. For continuous tags, such as consumption level, represented by average monthly consumption amount, a piecewise linear normalization method can be used to map them to the interval between 0 and 1, and then the fully connected layer is used to map them into a 128-dimensional vector.

[0015] Fusion of profile tags and spatial features: The encoded profile tag vector and scene feature vector, including the encoded vectors of travel time and mode of travel, are concatenated to form a comprehensive input feature vector. This vector is used as the input of the prediction model and participates in spatial attention calculation.

[0016] Through the above processing, the profile labels of travelers can be transformed from raw information annotations into computable numerical features, and then fused with spatial geographic features through nonlinear transformation of neural networks, ultimately affecting the output of the prediction model. This process has clear technical means and the effect of improving prediction accuracy. Among them, the technical means include embedding encoding, feature fusion, neural network calculation, etc.

[0017] It should be noted that when the personal information data involved in this application is used in specific products or technologies, user permission or consent will be obtained in advance, and the collection, use and processing of the relevant data will also comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0018] S20: Input the prediction constraints into the pre-trained prediction model to generate the prediction coordinate matrix and its corresponding selection probability matrix; In this embodiment, the prediction model refers to a computational model obtained after training, used to output prediction results based on input prediction constraints. The prediction model is primarily responsible for transforming abstract constraints into specific spatial coordinates and their selection probabilities. The prediction coordinate matrix is ​​a set of discrete geographic coordinate points output by the prediction model, where each coordinate point represents a corresponding potential travel destination. The selection probability matrix is ​​a set of probability values ​​corresponding one-to-one with each coordinate point in the prediction coordinate matrix, representing the likelihood of that coordinate point being selected as a travel destination. The selection probability matrix is ​​another output data structure corresponding one-to-one with the prediction coordinate matrix, storing the conditional probability value of each potential destination coordinate point being selected. The selection probability matrix, combined with the prediction coordinate matrix, comprehensively describes the output prediction results of the prediction model.

[0019] Specifically, the prediction constraints are input into the pre-trained prediction model to generate a prediction coordinate matrix and its corresponding selection probability matrix. In this step, the prediction model receives the above-mentioned prediction constraints and processes and analyzes them. As one implementation method, the prediction model can be a regression model based on historical statistical data. It learns the frequency patterns of users going to different locations under specific constraints in historical travel data, thereby predicting the coordinates of potential travel destinations and their probability of being selected. For example, after receiving the constraint "Tomorrow morning at 9 o'clock, leaving the company, walking, young white-collar worker", the prediction model finds by querying historical data that this type of person most often goes to "Café A", "Park B", and "Restaurant C" under the same time period and travel mode. The prediction model then directly outputs these three locations as prediction coordinates and uses the historical access frequency as the selection probability to form a prediction coordinate matrix and a selection probability matrix.

[0020] S30: Based on the predicted coordinate matrix and the selection probability matrix, a continuous probability distribution map of the target area information is generated through a predefined spatial interpolation strategy, and the continuous probability distribution map is rendered as a probability distribution heatmap in the interactive scene; In this embodiment, spatial interpolation strategy refers to an algorithmic method that converts discrete point data into a continuous spatial surface. It estimates the probability of any location within the entire spatial region using the probability values ​​of known points. Common methods include Kriging and inverse distance weighting. A continuous probability distribution map refers to a continuous and uninterrupted probability density surface in space. Specifically, the continuous probability distribution map is generated by calculating and filling the probability values ​​of discrete points using spatial interpolation techniques, ultimately covering the entire target analysis area. Furthermore, on the continuous probability distribution map, any geographical location has a corresponding probability value, indicating the likelihood that the location will become a destination. The probability of a trip being a destination is categorized into several aspects: probability distribution, ...

[0021] Specifically, based on the predicted coordinate matrix and the selection probability matrix, a continuous probability distribution map of the target area information is generated through a predefined spatial interpolation strategy. This continuous probability distribution map is then rendered as a probability distribution heatmap in an interactive scene. More specifically, the discrete coordinate points and their probability values ​​output by the prediction model are transformed into a continuous probability distribution covering the entire target area using a spatial interpolation algorithm. As an implementation method, the spatial interpolation strategy can employ simple linear interpolation or nearest-neighbor interpolation. For example, the target area can be divided into a grid of fixed size. For each grid cell, the distance to the nearest few points in the predicted coordinate matrix is ​​calculated, and a weighted average is taken based on the distance to obtain the probability value of that grid cell. Subsequently, the continuous probability distribution map is mapped to a color gradient and presented as a heatmap in an interactive map interface that allows users to intuitively perceive the travel probability of different areas.

[0022] S40: In response to the user's interactive selection operation of the target area in the probability distribution heatmap, determine the geometric boundary information of the target area; In this embodiment, geometric boundary information refers to polygonal boundary data that can be processed by a computer, generated based on the selection action and the transformation of the selected area after the user interactively selects a region on the probability distribution heatmap.

[0023] Specifically, in response to the user's interactive selection of a target area in the probability distribution heatmap, the geometric boundary information of the target area is determined. Specifically, when a user observes the probability distribution heatmap in an interactive scenario, they may interactively select a high-probability area or a specific area of ​​interest. As one implementation method, the user can directly select the area of ​​interest on the heatmap by clicking, dragging a rectangle, or drawing a polygon. After receiving the user's interactive operation, the screen coordinates corresponding to the interactive operation are converted into geographic coordinates, and the geometric boundary information of the selected area is generated, such as the coordinates of the four corner points of a rectangle or the vertex sequence of a polygon, for subsequent point of interest identification.

[0024] S50: Identify all first points of interest within the target area based on geometric boundary information, and obtain business distribution data associated with each first point of interest; In this embodiment, the first point of interest information refers to geographic entity data with specific functions or attractiveness within the target area, such as commercial facilities, public service facilities, cultural and entertainment venues, etc.; the business format distribution data refers to detailed information such as business type, service content, and scale associated with each first point of interest information, used to describe the business ecosystem characteristics of the first point of interest, such as whether a point of interest is classified as a "restaurant", "shopping center" or "park", as well as statistical information such as the number and proportion of each type of point of interest within the target area.

[0025] Specifically, based on geometric boundary information, all first points of interest (POIs) within the target area are identified, and the business distribution data associated with each POI is obtained. More specifically, after determining the geometric boundary information of the target area of ​​interest to which the user is interested, this information is used to retrieve all POIs located within that area from a pre-set geographic information database. As one implementation method, a pre-set POI database can be connected. Upon receiving the geometric boundary information, a spatial query is performed based on the POI database to filter out all POIs whose geographical locations fall within the geometric boundary. For each identified POI, its associated attribute data, such as name, address, and category, is further queried as business distribution data.

[0026] S60: Input the selection probability matrix, the first point of interest information and its associated business distribution data into the pre-trained semantic analysis model to generate a prediction conclusion report.

[0027] In this embodiment, the semantic analysis model refers to a computational model that has been trained and is capable of semantic understanding and reasoning on input structured or unstructured data. In this embodiment, the semantic analysis model is used to analyze the relationship between the prediction results and the actual points of interest, and generate a report with business insights. The prediction conclusion report is the final output of the entire travel demand prediction process. Specifically, it can be an analysis summary expressed in natural language, which intuitively summarizes the insights into the travel preferences of a specific group under specific conditions, so as to directly support decision-making.

[0028] Specifically, the selection probability matrix, information on the first point of interest, and related business distribution data are input into a pre-trained semantic analysis model to generate a prediction report. In other words, after receiving all the data, the semantic analysis model performs a comprehensive analysis. As one implementation method, the semantic analysis model can be a rule-based simple text generator. Upon receiving the selection probability matrix, information on the first point of interest, and business distribution data, it fills these data into a report template according to preset templates and keyword matching rules. For example, it can count the number and type of points of interest in high-probability areas and then generate descriptive text. Finally, the semantic analysis model outputs a structured or unstructured prediction report, providing users with analysis and decision-making suggestions regarding travel needs in the target area.

[0029] For example, suppose user A is a commercial real estate developer who wants to know which areas young families starting from residential area X are most likely to visit for leisure activities on a weekend afternoon, in order to make commercial site selections.

[0030] First, user A inputs a demand prediction command through the user interface, specifying the travel origin as "residential area X", the travel time as "weekend afternoon", the travel mode as "driving", and selecting the traveler profile label as "young family". After receiving the above input, based on the geographical location of residential area X, the target area is determined to be within a 10-kilometer radius. At the same time, based on the profile label "young family", the target group that meets this characteristic is filtered from the internal user database. Thus, the prediction constraints are constructed as follows: travel time: weekend afternoon, travel mode: driving, traveler profile label: young family.

[0031] Furthermore, the prediction constraints are input into the pre-trained prediction model, which, based on its internal prediction logic, learns the travel destination patterns of young families under similar constraints by analyzing a large amount of historical travel data. After processing, the prediction model outputs a set of discrete prediction coordinate matrices, such as containing the coordinates of potential destinations like "Shopping Mall A", "Park B", "Amusement Park C", and "Cinema D". At the same time, the prediction model also generates corresponding selection probability matrices, such as [0.4, 0.3, 0.2, 0.1], to indicate that under given conditions, young families are most likely to go to Shopping Mall A, followed by Park B, and so on.

[0032] Furthermore, based on the predicted coordinate matrix and the selection probability matrix, a predefined spatial interpolation strategy, such as distance-weighted interpolation, is used to transform the discrete prediction results into a continuous probability distribution map covering the entire target area. Subsequently, in the interactive scene, the continuous probability distribution map is rendered as a probability distribution heatmap. Thus, user A can observe on the map interface that the area centered on shopping mall A is the darkest color, indicating that this area is the area with the highest probability of young families traveling.

[0033] Furthermore, in response to user A dragging the mouse to select the high-probability area on the interactive interface, the geographic coordinates of the rectangular area are determined as the geometric boundary information of the target area based on this interactive selection operation.

[0034] Furthermore, based on this geometric boundary information, a spatial query is performed in the preset point of interest database to identify all primary points of interest within the area, such as: shopping mall A, children's playground E, family-friendly restaurant F, supermarket G, etc. At the same time, the business distribution data associated with these primary points of interest are obtained from the database. For example, shopping mall A is a comprehensive commercial complex that includes catering, retail, and entertainment; children's playground E is a children's entertainment facility; family-friendly restaurant F is a family-themed restaurant; and supermarket G is a convenience store.

[0035] Furthermore, the probability matrix, the first point of interest information, and the associated business distribution data are input into a pre-trained semantic analysis model. This model performs a comprehensive analysis of the multi-source data, for example, identifying the dominant business types in the high-probability area as "commercial complexes," "children's entertainment," and "family-friendly dining." Finally, the semantic analysis model generates a prediction report based on a preset report template, which may include: "According to the prediction, young families starting from residential area X are most likely to visit the area centered on shopping mall A on weekend afternoons. In this area, children's playgrounds, family-friendly restaurants, and other businesses are highly concentrated, indicating that this area has a strong attraction for young families. It is recommended to consider developing or introducing more family-friendly commercial facilities in the surrounding area." Compared to existing technical solutions that typically rely on macroscopic historical statistical data analysis, this embodiment generates a prediction coordinate matrix and a selection probability matrix, and further generates a continuous probability distribution map and a probability distribution heat map through a spatial interpolation strategy, thereby achieving a refined prediction of the spatial distribution of residents' travel destinations. In the example above, user A can intuitively understand that the area centered on shopping mall A presents the highest travel probability. This presentation is far more accurate than simply providing regional-level aggregated data, and can thus effectively support decisions such as business site selection, solving the problem of coarse spatial granularity in prediction results in existing technologies.

[0036] Furthermore, this embodiment introduces a probability distribution heatmap in an interactive scenario and supports users to interactively select target areas. In the example, user A can intuitively perceive the probability of travel in different areas and determine the geometric boundary information of the area of ​​interest through simple interactive operations based on their own focus. This allows users to dynamically perceive changes in potential destinations and conduct in-depth analysis of specific areas of interest, which can effectively make up for the shortcomings of existing technologies in interactive exploration capabilities.

[0037] More importantly, this embodiment combines data analysis, spatial visualization, and semantic interpretation. Specifically, by inputting the selection probability matrix, the first point of interest information, and its associated business distribution data into a pre-trained semantic analysis model, a predictive report with business insights is generated. In the example, the semantic analysis model not only identifies the main business types within high-probability areas but also proposes specific business development suggestions. This complete data-to-decision process effectively solves the problem of the separation between data analysis, spatial visualization, and semantic interpretation in existing technologies, thereby enhancing the application value of real-time decision support and providing users with easily understandable and business-supporting conclusive reports, thus improving the efficiency and practicality of decision-making.

[0038] In one embodiment, the prediction model includes a feature embedding layer, a spatial attention layer, and a probability output layer, such as... Figure 2 As shown, step S20 includes: S21: The feature embedding layer maps the received prediction constraints into corresponding dense feature vectors, and concatenates and nonlinearly fuses the dense feature vectors to generate scene feature vectors. In this embodiment, the feature embedding layer refers to the module layer in the prediction model, which is used to transform the original heterogeneous input data into a unified and continuous numerical representation, i.e., a dense feature vector. Specifically, its role is to capture the semantic information of the input features and provide rich context for the model. The implementation can include using a multilayer perceptron to independently encode different types of input features, or using the embedding layer in the Transformer encoder to process sequence data. The dense feature vector is a continuous numerical array of real numbers output by the feature embedding layer. Unlike the original independent classification labels, each dense feature vector occupies a point in a continuous vector space. The values ​​in each dimension of the vector represent a certain potential semantic feature, making similar concepts closer in the vector space. The scene feature vector is a single comprehensive vector formed by splicing and fusing multiple dense feature vectors such as travel time, mode, and crowd, which represents a specific travel scene set by the user.

[0039] Specifically, step S21 aims to unify the representation of various types of prediction constraints into a numerical form that the model can process, and fuse them to form a comprehensive scene description feature vector. Its function is to provide a unified input containing all relevant contextual information for the subsequent spatial attention mechanism. One possible implementation is to first use an independent embedding layer to generate a dense feature vector for each prediction constraint. For example, recurrent embedding can be used for time information, and word embedding or one-hot encoding followed by a multilayer perceptron can be used for image labels. The generated dense feature vectors are then concatenated and nonlinearly fused through one or more fully connected layers combined with activation functions to capture the complex interaction relationships between different features, thereby generating a scene feature vector. Another implementation is to use sequence models such as gated recurrent units or long short-term memory networks to gradually fuse different types of features to generate a scene feature vector.

[0040] S22: The spatial attention layer uses the origin information of the trip as the query center and the information of the second points of interest within a preset range around the query center as the key value to construct a spatial attention graph. It generates the spatial association weight of each second point of interest relative to the origin information of the trip by calculating the semantic association degree between the query center and the key value and the geographical decay factor. In this embodiment, the spatial attention layer refers to a module layer in the prediction model, used to simulate the attention mechanism for spatial information during decision-making. By calculating the degree of correlation between different spatial locations and the query center, it highlights spatial regions more relevant to the prediction target. Specifically, its role is to enhance the model's understanding of spatial context and provide weighted information for subsequent probability outputs. Its implementation can include construction based on self-attention or cross-attention mechanisms, where query, key, and value vectors can be generated by different linear transformations. The spatial attention graph refers to the intermediate product or visualization result of the spatial attention layer's calculation process, which can show the degree of attention paid to different geographical locations around the starting point. Semantic association refers to the quantitative index used in the spatial attention mechanism to measure the degree of functional or semantic matching between a second point of interest and the current travel scenario. The geographic decay factor is a weighting coefficient that decreases with increasing geographic distance, reflecting the "distance friction" effect. This means that even if a location is semantically highly relevant, its likelihood of being selected decreases due to travel costs if it is too far from the starting point. The geographic decay factor ensures the rationality of spatial relationships. Second point of interest information refers to the set of all candidate first points of interest (POIs) retrieved from the geographic information database within a preset search radius centered on the user's previously selected starting point during the prediction model's calculation process. Spatial association weight is the output of the spatial attention layer, specifically the combined score of semantic relevance and the geographic decay factor. It represents the final relative attractiveness of each second point of interest relative to the starting point; the higher the weight, the greater the likelihood of that second point of interest being selected.

[0041] Specifically, step S22 aims to dynamically evaluate the attractiveness or relevance of different second points of interest to given travel origin information through an attention mechanism. Its role is to provide the model with a mechanism that enables it to intelligently focus on the most relevant potential destinations based on the current travel context. This can be achieved by: firstly, encoding the travel origin information as a query vector and encoding the second point of interest information within a preset range as key and value vectors; calculating the similarity between the query vector and each key vector as semantic relevance, for example, using dot product or cosine similarity; simultaneously, calculating a geographical attenuation factor based on the geographical distance between the travel origin information and the second points of interest, for example, the greater the distance, the smaller the attenuation factor, which can be calculated using an exponential attenuation function or a Gaussian attenuation function; finally, weighting the semantic relevance and the geographical attenuation factor and normalizing it using a Softmax function to obtain the spatial relevance weight of each second point of interest relative to the travel origin information; or by constructing a spatial attention graph using a graph neural network, where nodes represent second points of interest and edges represent spatial relationships between nodes, and calculating spatial relevance weights through a message passing mechanism.

[0042] Furthermore, as an example, the geographic decay factor can be expressed as an exponential decay function, specifically as follows: ,in, This is the geographical distance between the origin of travel and the second point of interest, in meters, which can be calculated using the Haversine formula. This is the maximum search radius for the target area information, used for distance normalization. This is the decay coefficient, which can be set as a learnable parameter; during training, The initial value can be set to 1.0. After automatic optimization through backpropagation, its final convergence value is usually between 0.5 and 2.0.

[0043] Furthermore, as another example, the geographic attenuation factor can also be expressed as a power-law attenuation function, with the following expression: ,in, The feature distance parameter represents the distance when the distance is... At that time, the decay factor drops to half of its initial value. The value can be dynamically adjusted according to the mode of travel: when walking It can be set to equal 300 meters when cycling. It can be set to 1500 meters when driving. It can be set to 5000 meters, etc.; The attenuation index can be set from 1.5 to 3.0, with a preferred value of 2.0. The larger the value, the more significant the impact of distance on weight.

[0044] By introducing a geographic attenuation factor, the spatial attention layer can further consider travelers' sensitivity to distance on the basis of calculating semantic relevance, making the prediction results more consistent with the distance attenuation law, thereby improving prediction accuracy.

[0045] S23: The probability output layer sorts and filters the second interest points based on spatial correlation weights, performs probability normalization on the spatial correlation weights of the filtered second interest point set, and outputs the prediction coordinate matrix and its corresponding selection probability moments.

[0046] In this embodiment, the probability output layer refers to the module layer in the prediction model, which is used to convert the spatial association weights generated by the spatial attention layer into an interpretable probability distribution, namely the prediction coordinate matrix and its corresponding selection probability matrix. Specifically, its function is to quantify the attractiveness or probability of selection of each potential travel destination. The implementation method may include using the Softmax function to convert the normalized weights into a probability distribution, or using a Gaussian mixture model to fit the distribution of potential travel destinations. The probability normalization process refers to transforming the original spatial association weights of different magnitudes output by the spatial attention layer into a probability distribution that conforms to the probability axiom through mathematical transformation. After the probability normalization process, the sum of all weights is 1, and each weight value is the probability of the corresponding second point of interest being selected, thus obtaining the selection probability matrix.

[0047] Specifically, step S23 aims to transform the weights representing the importance of each second point of interest generated by the spatial attention layer into specific travel destination prediction results. Its function is to provide quantifiable travel demand predictions, including possible destination coordinates and the probability of selecting these destinations. This can be achieved by first sorting all second points of interest in descending order according to spatial association weights, and selecting the N points of interest with the highest weights as potential destinations to form a prediction coordinate matrix, i.e., containing the geographical coordinates of these second points of interest. Then, the spatial association weights of these N selected points of interest are subjected to Softmax normalization to ensure that the sum of all probabilities is 1, thereby obtaining the corresponding selection probability matrix. Another implementation method is that the probability output layer can directly output a multinomial distribution parameter, where each category of the multinomial distribution corresponds to a second point of interest, and the prediction result is generated by sampling or directly selecting the point with the highest probability.

[0048] Specifically, the proposed solution constructs a prediction model consisting of a feature embedding layer, a spatial attention layer, and a probability output layer to achieve refined processing of residents' travel demand predictions. Specifically, the feature embedding layer receives and processes diverse prediction constraints, transforming them into a unified scene feature vector rich in semantic information. The spatial attention layer, centered on the travel origin information and combined with surrounding secondary points of interest (POIs), constructs a dynamic spatial attention mechanism. This mechanism not only considers the semantic correlation between the travel origin and potential destination but also introduces a geographical attenuation factor, thus more realistically reflecting the traveler's sensitivity to distance. In this way, the most relevant secondary POIs to the current travel context are identified, and corresponding spatial correlation weights are assigned to them. Finally, the probability output layer sorts and filters the secondary POIs based on the spatial correlation weights and performs probability normalization processing to generate a prediction coordinate matrix and a selection probability matrix.

[0049] Furthermore, as an example, the training process of the prediction model is as follows: Training data construction: Obtain anonymized historical travel trajectory data from entities such as urban traffic management departments, with a data time span of no less than 6 months and a total sample size of no less than 1 million. Each sample includes the coordinates of the origin and destination, the travel time accurate to the hour, the mode of travel (walking, cycling, driving, public transportation), and user profile tags including age group, occupation type, and consumption level. Spatial matching is performed between the destination coordinates and a pre-set point of interest database, with a matching threshold of 50 meters. Each trajectory is labeled with its corresponding destination point of interest ID and category.

[0050] Feature embedding layer training: For discrete features, including travel time, travel mode, and population profile labels, a learnable embedding matrix can be used to map them into a 128-dimensional dense vector. The embedding matrix is ​​jointly trained with the overall model through backpropagation.

[0051] Spatial attention layer training: The core parameters of the spatial attention layer include the query matrix, key matrix, and value matrix, all of which are 128 by 128. The geographic decay factor adopts a learnable parameterized form, and the initial value of its learnable parameter can be set to 0.0001, which is automatically optimized during training through gradient descent.

[0052] Loss function design: Cross-entropy loss function or other methods can be used to measure the difference between the predicted choice probability matrix and the actual destination.

[0053] Training strategy: The Adam optimizer can be used, with a learning rate of 0.001, a batch size of 256, and 50 training epochs. The loss function value is monitored on the validation set, and training is stopped early when the loss no longer decreases after 5 consecutive validation epochs to prevent overfitting.

[0054] Model Evaluation: The model performance was evaluated on an independent test set, using Top-5 accuracy as the evaluation metric, which is the proportion of the top 5 points of interest with the highest predicted probability that contain the real destination.

[0055] It should be noted that those skilled in the art can reproduce and train the prediction model based on the examples and content described in this application specification, combined with actual needs.

[0056] Through the above technical solutions, this application clarifies the internal structure and workflow of the prediction model, enabling it to more effectively handle multi-source heterogeneous prediction constraints. Specifically, the feature embedding layer transforms complex heterogeneous information into a unified scene feature vector; the spatial attention layer, by introducing semantic relevance and geographical attenuation factors, fully considers travelers' preferences for spatial location and their sensitivity to distance during the prediction process, thereby more accurately capturing potential travel destinations; the probability output layer transforms the spatial relevance information generated by the spatial attention layer into an intuitive prediction coordinate matrix and selection probability matrix, thereby improving the accuracy and interpretability of the prediction results. In summary, this application, through the above-described prediction model design architecture, can effectively solve the problems of insufficient prediction accuracy and poor model interpretability that may exist in traditional prediction methods when dealing with complex spatial dependencies and personalized user needs, providing more reliable data support for predicting residents' travel demand.

[0057] In one embodiment, step S21 includes: S211: Map travel time information into periodic temporal embedding vectors and map traveler profile tags into semantic embedding vectors; In this embodiment, travel time information is mapped to a periodic temporal embedding vector. This aims to transform discrete or continuous time information into a low-dimensional vector representation while preserving its periodicity, thus understanding the impact of time on travel demand, such as morning and evening rush hours and weekend effects. For example, travel time information can be mapped to a two-dimensional space using sine and cosine functions to capture its periodicity; or, for discrete time units, one-hot encoding can be performed first, and then an embedding layer can be used to convert it into a dense vector; or, for continuous timestamps, they can be decomposed into components such as year, month, day, hour, and minute, and each component can be embedded or encoded separately before being concatenated. Mapping travel profile labels to semantic embedding vectors aims to transform these discrete or categorical profile labels into low-dimensional vector representations, making semantically similar labels closer together in the vector space. This helps to understand the travel preferences of different groups. For example, for categorical labels such as occupation and family structure, one-hot encoding can be performed first, and then an embedding layer can be used to convert them into dense vectors. Alternatively, for textual descriptive labels, a pre-trained word embedding model can be used to convert the text description into a semantic vector. Or, multiple discrete or continuous profile features can be concatenated and then nonlinearly transformed through one or more fully connected layers to generate semantic embedding vectors.

[0058] S212: Based on travel mode information and origin information, calculate the distance decay curve that can be reached within a preset time period and encode it as a spatial reachability embedding vector. In this embodiment, step S212 aims to capture the impact of travel origin information and travel mode information on potential destination selection, reflecting the user's time and space constraints, thereby more accurately predicting the travel destination. For example, based on travel origin information, travel mode information, and preset duration, an reachable geographical area, such as an isochronous circle or an isodistance circle, can be calculated, and then the geometric features of the geographical area or the distribution features of its internal points of interest can be encoded into a vector; or, road network data from a geographic information system can be used to calculate the actual travel time or distance from the travel origin information to various surrounding points, and then a distance decay function can be applied to convert these times or distances into reachability scores, and the reachability scores can be encoded into vectors; or, the target area can be divided into grids, and for each grid cell, its reachability score from the travel origin information can be calculated, and the matrix or vector formed by these reachability scores can be used as a spatial reachability embedding vector.

[0059] S213: Concatenate the temporal periodic embedding vector, semantic embedding vector, and spatial reachability embedding vector into a scene feature vector.

[0060] In this embodiment, step S213 aims to fuse embedding vectors from different modalities into a unified high-dimensional scene feature vector through a concatenation operation. This scene feature vector comprehensively represents the user's overall needs and environmental characteristics at a specific time, location, and mode of travel. For example, the three embedding vectors from the previous step can be directly connected in a predetermined order to form a new vector with a dimension equal to the sum of the dimensions of the three vectors; or, before concatenation, learnable weights can be assigned to each embedding vector to adjust the relative importance of these embedding vectors in the final scene feature vector; or, after concatenation, one or more fully connected layers can be used to perform a nonlinear transformation on the concatenated vector to capture the complex interaction relationships between different features and further enhance the feature representation capability.

[0061] Specifically, to comprehensively and effectively capture the impact of heterogeneous input information on residents' travel demand, the feature embedding layer first performs embedding processing on different types of constraints. Specifically, for travel time information with periodic patterns, mapping it to periodic temporal embedding vectors encodes the periodic features in the time dimension into a low-dimensional vector space, enabling the prediction model to identify and utilize the periodic impact of time on travel behavior, such as distinguishing travel differences between weekdays and weekends, and between peak and off-peak hours. Simultaneously, for travel population profile tags describing individual user characteristics, mapping them to semantic embedding vectors transforms categorical population features into dense semantic representations, allowing people with similar travel preferences to be represented in the vector space. By recognizing proximity in space, the system captures the personalized needs of different user groups. Furthermore, considering the spatial constraints and accessibility of travel behavior, it calculates the distance decay curve reachable within a preset time period based on travel mode and origin information, and encodes it as a spatial accessibility embedding vector. This quantifies the accessibility and attractiveness of different geographical locations from the origin under specific travel modes and time budgets, effectively reflecting the constraints and guidance of spatial factors on travel destination selection. The system then concatenates these temporal periodic embedding vectors, semantic embedding vectors, and spatial accessibility embedding vectors extracted from time, semantic, and spatial dimensions, respectively, to integrate multimodal feature information into a unified high-dimensional scene feature vector.

[0062] Through the above technical solutions, this application can transform heterogeneous and complexly related prediction constraints, including travel time information, traveler profile tags, travel mode information, and travel origin information, into unified and semantically rich scene feature vectors. Specifically, the introduction of periodic temporal embedding vectors can accurately capture periodic patterns in the time dimension, avoiding information loss that may result from simple time encoding. The generation of semantic embedding vectors allows for a detailed characterization of the personalized travel preferences of different groups, enhancing the prediction model's ability to understand user profiles. The encoding of spatial accessibility embedding vectors quantifies the limitations of travel origin and mode on spatial selection, enabling the prediction model to more realistically reflect users' spatial behavior patterns. Through this multi-dimensional feature embedding and fusion mechanism, the expressive power of scene feature vectors can be improved, thereby enhancing the accuracy and robustness of predicting residents' travel demand.

[0063] In one embodiment, step S22 includes: S221: Based on the scene feature vector, a query vector representing the travel origin information, a key vector representing the information of each second point of interest, and a value vector for carrying semantic information are generated through independent linear transformation layers. In this embodiment, an independent linear transformation layer refers to an independent computational unit in the neural network model that performs linear mapping on the input data. Its function is to project the data from the original feature space onto a new feature space, thereby extracting feature representations of different dimensions or semantics. For example, fully connected layers or convolutional layers can be used as linear transformation layers. By learning weight matrices and bias terms, the input scene feature vector is converted into query vectors, key vectors, and value vectors. Furthermore, these linear transformation layers are independent of each other, meaning that they have their own independent parameters and can learn different feature mapping relationships for different output vectors. The query vector is a feature representation of the query center, which is used in the attention mechanism to match with the key vector to determine relevance. The key vector is a feature representation of the object to be matched, which is used in the attention mechanism to compare with the query vector to calculate attention weights. The value vector is a feature representation that carries semantic information and usually corresponds to the key vector. After calculating the attention weights, the value vectors are weighted and summed to aggregate relevant information.

[0064] S222: The initial semantic association between each second point of interest and the origin information is obtained by calculating the cosine similarity between the query vector and each key vector; In this embodiment, cosine similarity is a metric that measures the cosine of the angle between two non-zero vectors. It is often used to evaluate the directional similarity between vectors. Its value ranges from -1 to 1. The closer the value is to 1, the more consistent the directions of the two vectors are and the higher the similarity. Initial semantic association refers to the degree of association between the travel origin information and the information of each second point of interest obtained based solely on semantic information before considering geographical factors.

[0065] S223: Obtain the spherical distance between the origin of travel and each second point of interest, and generate the geographic attenuation factor corresponding to the spherical distance through a preset attenuation function; In this embodiment, spherical distance refers to the shortest distance between two points on the Earth's surface along a great circle arc. Since the Earth is an approximate sphere, spherical distance more accurately reflects actual geographical distance than Euclidean distance. The preset attenuation function is a mathematical function used to describe the law that a certain effect weakens as distance increases. In this embodiment, it is used to describe geographical correlation. Common attenuation functions include exponential attenuation function, linear attenuation function, or Gaussian attenuation function. The geographical attenuation factor is a weighting factor calculated by using the preset attenuation function based on the spherical distance between the travel origin information and each second point of interest. It reflects the negative impact of geographical distance on travel choices.

[0066] S224: Weight the initial semantic correlation degree of each second point of interest with its corresponding geographic decay factor, and perform fusion and normalization processing on all second points of interest to generate the spatial correlation weight of each second point of interest relative to the travel origin information.

[0067] In this embodiment, weighting refers to combining or fusing different values ​​according to a certain weight ratio. In this embodiment, the objects are the initial semantic relevance and the geographic decay factor. Fusion normalization processing refers to standardizing the result after weighting and fusing multiple values ​​so that it falls within a specific range, and the sum of all processed values ​​is 1. Common normalization methods include the Softmax function. Spatial association weight refers to the comprehensive importance or attractiveness index of each second point of interest relative to the travel origin information after comprehensively considering semantic relevance and geographic decay factor.

[0068] Specifically, based on scene feature vectors, independent linear transformation layers can be used to generate query vectors representing travel origin information, key vectors representing each second point of interest (POI), and value vectors carrying semantic information. This transforms abstract scene features into specific representations required by the attention mechanism, enabling the prediction model to understand travel origins and potential destinations from different perspectives. By calculating the cosine similarity between the query vector and each key vector, the initial semantic correlation between each POI and the travel origin information is obtained, aiming to quantify the semantic matching degree between the travel origin and each POI, reflecting the user's potential interest preferences. Simultaneously, the spherical distance between the travel origin information and each POI is obtained, and a geographic attenuation factor corresponding to the spherical distance is generated through a preset attenuation function. This geographic attenuation factor can intuitively reflect the negative impact of geographic distance on travel choices; the greater the distance, the more significant the attenuation. Finally, the initial semantic correlation of each POI is weighted by its corresponding geographic attenuation factor, and all POIs are fused and normalized to generate the spatial correlation weight of each POI relative to the travel origin information.

[0069] Through the above technical solution, this application comprehensively considers the dual influence of semantic relevance and geographical distance when generating spatial association weights for each second point of interest relative to the travel origin information. Specifically, the scene feature vector is decomposed into query vector, key vector, and value vector, enabling the prediction model to capture the association between the travel origin and potential destination from a semantic level. At the same time, spherical distance and a preset decay function are introduced to generate a geographical decay factor, which can effectively quantify the objective limitations of geographical distance on travel decisions. Furthermore, the semantic relevance and geographical decay factor are weighted, fused, and normalized, so that the final generated spatial association weights not only reflect users' interest preferences for destinations but also fully consider actual geographical accessibility. This improves the accuracy and rationality of the prediction model in generating prediction coordinate matrices and selection probability matrices, providing more reliable data support for urban planning and business layout.

[0070] In one embodiment, step S30 includes: S31: The grid division granularity is determined based on the spatial distribution density of points in the predicted coordinate matrix by using a pre-set dynamic granularity determination strategy, and several grid cells are generated according to the grid division granularity. In this embodiment, the pre-defined dynamic granularity determination strategy refers to a pre-defined set of rules or algorithms used to adjust the spatial grid division precision according to specific conditions. Its core lies in selecting an appropriate grid size based on the characteristics of the input data to avoid over-refinement in sparse data regions or over-coarsening in dense data regions. In other words, this dynamic granularity determination strategy aims to optimize the efficiency and accuracy of spatial interpolation. One implementation is to define a series of thresholds; when the spatial distribution density of points is below a certain threshold, a coarse-grained grid is used, and when it is above another threshold, a fine-grained grid is used. Another implementation is to use a method based on spatial index structures such as quadtrees or Kd-trees to adaptively divide the space according to the distribution of data points, thereby obtaining grids of different granularities. Determining the grid division granularity based on the spatial distribution density of points in the predicted coordinate matrix aims to select the most suitable grid size based on the actual distribution of discrete points in the predicted coordinate matrix. Its purpose is to ensure that sufficiently high resolution is provided in dense data regions to capture details, while a lower resolution is used in sparse data regions. Resolution is improved to reduce computational load and visual noise. One approach is to calculate the number of points in a local area or the reciprocal of the distance between points as a density index, and then associate this density index with a preset granularity mapping function to obtain the corresponding grid granularity. Another approach is to use kernel density estimation to estimate the spatial distribution density and dynamically adjust the grid size based on the estimation results; for example, the higher the density, the finer the grid granularity. Generating several grid cells based on the grid granularity is to build a basic spatial discretization structure for subsequent spatial interpolation operations. By dividing the target area into a series of regular or irregular grid cells according to the previously determined grid granularity, each grid cell represents a small spatial region. One approach is to use uniform grid partitioning, that is, to use a uniform grid granularity to partition the entire target area. Another approach is to use non-uniform grid partitioning, for example, to generate smaller grid cells in high-density areas and larger grid cells in low-density areas, forming an adaptive grid structure.

[0071] S32: Using the Gaussian kernel function interpolation algorithm, spatial interpolation is performed on the predicted coordinate matrix and the selection probability matrix on the generated grid cells to generate a continuous probability distribution map; In this embodiment, the Gaussian kernel interpolation algorithm is a commonly used spatial interpolation method. It estimates the value of unknown points by weighting the known data points. The Gaussian kernel function, as a weighting function, is characterized by the fact that the closer the interpolation point is to the observation point, the greater its weight, and the weight decreases with increasing distance in a Gaussian distribution. It is used to convert discrete prediction points and their probability values ​​into a continuous probability distribution. One implementation is to calculate the distance between the center point of each grid cell and all points in the prediction coordinate matrix, calculate the weight according to the Gaussian kernel function, and then apply the calculated weight to the probability values ​​in the corresponding selection probability matrix for weighted summation. Another implementation is to use an optimized Gaussian kernel interpolation algorithm, for example, by using spatial index structures such as KD trees or R trees to accelerate the search for nearest neighbors, thereby improving interpolation efficiency. Spatial interpolation processing of the prediction coordinate matrix and selection probability matrix on the generated grid cells refers to using an interpolation algorithm to convert discrete prediction results into continuous spatial distribution data, and to interpolate the point positions in the prediction coordinate matrix and the probability values ​​in the selection probability matrix. As input, the probability value of the center point or representative point of each grid cell is calculated on the pre-generated grid cells. One implementation is to use the center point of each grid cell as the interpolation point and calculate the probability value of that point using the Gaussian kernel function interpolation algorithm. Another implementation is to perform a weighted average of all predicted points within each grid cell, with the weight determined by the distance from the predicted point to the grid center point, thus obtaining the average probability value of that grid cell. A continuous probability distribution map refers to a two-dimensional or three-dimensional data structure in which each spatial location or grid cell is associated with a predicted probability value. These probability values ​​change continuously in space and can reflect the smooth distribution trend of travel demand within the target area. The continuous probability distribution map is the basic data for heatmap rendering. One implementation is to store the probability value of each grid cell obtained by interpolation in a two-dimensional array or matrix to form a raster data structure. Another implementation is to generate a vector data structure, such as a contour map, where each line connects points with the same probability value, thus visually representing the continuous change of probability.

[0072] S33: Associate the continuous probability distribution map with the real-time viewpoint height of the interactive scene, and dynamically adjust the rendering granularity and color transparency to generate a probability distribution heatmap.

[0073] In this embodiment, a continuous probability distribution map is associated with the real-time viewpoint height of the interactive scene. This aims to enable the heatmap's rendering effect to adaptively adjust according to the user's changing perspective within the interactive scene, providing a more human-perceptual visualization experience: when the user zooms in, the heatmap should display more detail; when the user zooms out, the heatmap should be more generalized. One implementation method is to obtain the current camera's Z-axis coordinates, i.e., the viewpoint height, in the rendering engine and input it as an input parameter into the rendering logic to calculate rendering granularity and color transparency. Another implementation method is to establish a mapping table or function between viewpoint height and rendering parameters, querying or calculating the corresponding rendering parameters when the viewpoint height changes. Dynamically adjusting rendering granularity and color transparency is key to achieving adaptive heatmap rendering. Rendering granularity determines the heatmap's level of detail, and color transparency determines the visual overlay effect. Dynamic adjustment ensures that better or optimal visual effects and information density are obtained at different viewpoints. One implementation method is... When the viewpoint height decreases, the rendering granularity is increased to display more details, and color transparency may be reduced to highlight local hotspots; when the viewpoint height increases, the rendering granularity is decreased to simplify the display, and color transparency may be increased to show the overall trend. Another implementation is to use multi-level LOD technology to pre-generate heatmap data of different granularities and dynamically switch the loading and rendering of data at different LOD levels according to the viewpoint height. A probability distribution heatmap is a visual chart that uses color depth or brightness to represent the density or magnitude of data within a spatial area. In this embodiment, the probability distribution heatmap intuitively displays the spatial distribution heat of residents' travel demand within the target area. One implementation is to use a graphics rendering library to map the probability values ​​in the continuous probability distribution map to a preset color gradient scheme, for example, from cool colors to warm colors, and render it as a texture or geometry in an interactive scene. Another implementation is to use pixel shader-based technology to calculate and render the heatmap in real time on the GPU, thereby achieving efficient dynamic visualization.

[0074] Specifically, upon receiving the prediction coordinate matrix and the selection probability matrix, to transform the discrete prediction results into an intuitive and continuous spatial distribution representation, a pre-set dynamic granularity determination strategy is first used. Based on the spatial distribution density of points in the prediction coordinate matrix, the grid division granularity is determined to ensure that a finer grid is used in densely populated areas of travel demand prediction points to capture local details, while a coarser grid is used in sparse areas to reduce unnecessary computational overhead and visual redundancy. A series of grid cells are generated according to the determined grid division granularity, providing a discretized spatial basis for subsequent spatial interpolation operations. Based on this, a Gaussian kernel function interpolation algorithm is used to perform positional information in the prediction coordinate matrix and probability values ​​in the selection probability matrix on the generated grid cells. Spatial interpolation using the Gaussian kernel function effectively diffuses discrete probability values ​​smoothly across the entire space, allowing the interpolation results to reflect continuous probability distribution trends and thus generating a continuous probability distribution map that associates information about the target area. To further enhance user experience and visualization, the generated continuous probability distribution map is correlated with the real-time viewpoint height of the interactive scene. This means that as the user zooms in or pans the viewpoint in the interactive scene, the rendering granularity and color transparency of the heatmap can be dynamically adjusted. For example, when the user zooms in, the heatmap is rendered with finer granularity and the color transparency may be adjusted to highlight local hotspots; when the user zooms out, the heatmap is rendered with more generalized granularity and the transparency may be increased to show the overall trend.

[0075] Through the above technical solutions, this application can transform discrete resident travel demand prediction results into intuitive, continuous, and adaptive probability distribution heatmaps. The dynamic granularity determination strategy combined with the Gaussian kernel interpolation algorithm effectively solves the sparsity or over-density problems that may occur in spatial visualization of discrete data, ensuring that the heatmap can display the distribution of travel demand with appropriate accuracy in different areas. Simultaneously, by highly correlated with the continuous probability distribution map and the real-time viewpoint of the interactive scene, and by dynamically adjusting the rendering granularity and color transparency, the user's perception and understanding efficiency of the spatial distribution of travel demand at different zoom levels can be improved. This allows users to clearly identify travel hotspots and perceive continuous changes in travel demand, thereby providing more accurate and intuitive decision support for urban planning.

[0076] In one embodiment, step S31 includes: S311: Calculate the spatial distribution density of all coordinate points in the predicted coordinate matrix within the target area, and identify high-density clustered areas and low-density sparse areas. In this embodiment, step S311 aims to quantify the density of predicted points within the target area, providing a basis for subsequent adaptive grid partitioning. By identifying regions with different densities, different processing strategies can be applied in a targeted manner. Specifically, a kernel density estimation method can be used to calculate the spatial distribution density of points. Kernel density estimation estimates a continuous density surface by placing a kernel function, such as a Gaussian kernel, around each data point and then superimposing all kernel functions. By setting a density threshold, regions with a density higher than the threshold can be identified as high-density clustered regions, and regions with a density lower than the threshold can be identified as low-density sparse regions. In addition, spatial index structures such as quadtrees or Kd-trees can be used to organize the predicted coordinate points. By traversing these tree structures, the number of points in each node or its coverage area can be counted, thereby indirectly reflecting the local point density.

[0077] S312: Apply a corresponding fine-grained grid division strategy to high-density clustered regions and a corresponding coarse-grained grid division strategy to low-density sparse regions to determine a multi-scale grid division scheme. In this embodiment, step S312 aims to dynamically adjust the mesh fineness based on the different density regions identified in the previous step, ensuring that more details are retained in critical regions while reducing computation in non-critical regions. The mesh granularity is negatively correlated with the spatial distribution density of points; that is, the denser the points, the finer the mesh; the sparser the points, the coarser the mesh. For example, a series of mesh granularity levels can be preset, such as 1 meter, 5 meters, 10 meters, and 50 meters. For identified high-density clustered regions, a smaller mesh granularity, such as 1 meter or 5 meters, is selected; for low-density sparse regions, a larger mesh granularity, such as 10 meters or 50 meters, is selected. Furthermore, this mapping relationship can be stored in a preset lookup table or defined through a preset piecewise function.

[0078] S313: Generate an adaptive mesh that includes mesh units of different granularities based on a multi-scale meshing scheme.

[0079] In this embodiment, step S313 aims to actually construct the final non-uniform grid structure for spatial interpolation based on the previously determined multi-scale partitioning scheme. This adaptive grid structure can more effectively capture data features and improve interpolation accuracy and efficiency. For example, an adaptive grid can be generated by recursive subdivision, starting from a coarse grid covering the entire target area. For grid cells identified as high-density clustered areas, they are further subdivided into smaller sub-grid cells until the preset minimum granularity is reached or the density requirement is met. Alternatively, irregular grid structures such as irregular triangular meshes or Voronoi diagrams can be used. In areas with high point density, the side length of the triangles in the irregular triangular mesh can be shorter, and the Voronoi polygons can be smaller, thereby achieving local refinement.

[0080] Specifically, before generating the continuous probability distribution map, the scheme in this application first performs a refined analysis of the spatial distribution of points in the predicted coordinate matrix. Specifically, by calculating the spatial distribution density of all predicted coordinate points within the target area, it can accurately identify high-density clustered regions with dense data points and low-density sparse regions with sparse data points. Based on this difference in density distribution, this application no longer uses a single grid granularity, but instead employs a finer-grained grid partitioning strategy for high-density clustered regions to capture local details; simultaneously, a coarser-grained grid partitioning strategy is used for low-density sparse regions to reduce unnecessary computational overhead. This design, where grid granularity is negatively correlated with point spatial distribution density, ensures high-resolution interpolation results in areas with abundant data information, while effectively saving computational resources in areas with limited data information. Finally, based on this multi-scale grid partitioning scheme, an adaptive grid containing grid units of different granularities is generated. This adaptive grid can better match the actual distribution characteristics of the predicted coordinate points, providing an efficient and accurate base grid for subsequent spatial interpolation processing. This enables the generated continuous probability distribution map to more realistically reflect the probability distribution of residents' travel needs and improves the efficiency and visual effect of heat map rendering.

[0081] Through the above technical solution, this application can dynamically adjust the granularity of the grid division according to the actual density of the spatial distribution of points in the predicted coordinate matrix. This allows for the use of a finer grid for analysis in areas where travel demand prediction points are highly concentrated, thereby more accurately capturing subtle changes in local travel demand and avoiding information loss due to an overly coarse grid. Conversely, in areas where travel demand prediction points are sparse, a coarser grid is used, effectively reducing unnecessary computation and storage overhead. This adaptive grid division strategy significantly improves the accuracy and efficiency of generating continuous probability distribution maps, enabling the final rendered probability distribution heatmap to more realistically and precisely reflect the actual spatial distribution of residents' travel demand, thus providing users with more accurate and valuable prediction conclusions.

[0082] Further, in one embodiment, step S32 includes: S321: Determine the bandwidth parameter of the Gaussian kernel function based on the spatial distribution characteristics of points in the predicted coordinate matrix; In this embodiment, the bandwidth parameter of the Gaussian kernel function determines the range and intensity of the influence of each observation point on the surrounding area during the interpolation process. Specifically, a smaller bandwidth will result in an interpolation result that is closer to the original data points, but may produce more local fluctuations; a larger bandwidth will result in a smoother interpolation result, but may lose local details. The bandwidth parameter can be evaluated on the training data using cross-validation to assess the interpolation error under different bandwidth parameters and select the bandwidth that minimizes the error. Alternatively, a data-driven approach can be used, such as the K-nearest neighbor algorithm, to dynamically estimate the local optimal bandwidth based on the distance distribution between each point in the predicted coordinate matrix and its nearest neighbor.

[0083] S322: Take the center point of each grid cell as the interpolation point and the coordinate point in the predicted coordinate matrix as the observation point. Calculate the spatial influence weight of each observation point on the interpolation point based on the Gaussian kernel function. In this embodiment, step S322 aims to quantify the contribution of each actual predicted point in the predicted coordinate matrix to the center of the grid cell; wherein, the Gaussian kernel function can assign a larger influence weight to the observation point closer to the observation point and a smaller influence weight to the observation point farther away, thereby achieving smooth spatial interpolation; for example, the Gaussian kernel function can be defined as: ,in, This is the value of the Gaussian kernel function, i.e., the calculated spatial influence weight, which ranges from 0 to 1. The weight is maximum (1) when the observation point coincides with the interpolation point, and decreases with distance. As the weight increases, the weight value monotonically decreases and approaches 0; It is the Euclidean distance between the observation point and the point to be interpolated. It is a predetermined bandwidth parameter used to control the rate at which the weights decay with distance.

[0084] Furthermore, other forms of kernel functions can also be used, such as the exponential kernel function or the Matern kernel function, which can also calculate the influence weights based on spatial distance, but have different decay characteristics.

[0085] S323: Based on the spatial influence weight, the selection probabilities corresponding to each observation point are weighted and fused, and the interpolation results and uncertainty estimates of the interpolation points to be interpolated are calculated and generated. In this embodiment, after obtaining the spatial influence weight of each observation point on the interpolation point, the spatial influence weight needs to be applied to the selection probability of each observation point to obtain the predicted probability value of the interpolation point. At the same time, to improve the reliability of the prediction, the uncertainty of the interpolation result also needs to be quantified, which helps users understand the confidence level of the prediction result. Among them, weighted fusion can be achieved by multiplying the selection probability of each observation point with its corresponding spatial influence weight, and then summing and normalizing all products. Uncertainty estimation can be obtained by calculating the variance or standard deviation of the weighted average. As another implementation method, Bayesian interpolation method can also be used, treating the selection probability of the observation point as a random variable, and constructing a posterior probability distribution in combination with the spatial influence weight, and extracting the mean and variance of the interpolation result as uncertainty estimation.

[0086] S324: Integrate the interpolation results and uncertainty estimates of each grid cell to generate a continuous probability distribution map that is associated with probability density and uncertainty information.

[0087] In this embodiment, after completing the interpolation calculation for all grid cells, the discrete interpolation results and uncertainty estimates need to be aggregated to form a complete and continuous probability distribution map. The continuous probability distribution map not only shows the probability density of residents' travel demand, but also provides information about the reliability of the predicted values. Furthermore, the interpolation result of each grid cell can be used as the probability density value of that grid cell, and the uncertainty estimate can be used as the confidence interval or error range of that grid cell. Then, the above information can be presented on a continuous geographic space using visualization tools in terms of color depth, transparency, or additional layers.

[0088] Specifically, to ensure the accuracy and adaptability of the interpolation process, the bandwidth parameter of the Gaussian kernel function is dynamically determined based on the spatial distribution characteristics of points in the predicted coordinate matrix. By adjusting the bandwidth according to the distribution characteristics of the data itself, over-smoothing or under-smoothing problems that may occur with a fixed bandwidth can be avoided, thus making the interpolation results more realistically reflect the potential distribution of residents' travel demand. The center points of each pre-divided grid cell are regarded as the interpolation points, while the actual predicted points in the predicted coordinate matrix are regarded as observation points. Using the determined Gaussian kernel function, the spatial influence weight of each observation point on the interpolation point is calculated. This distance-attenuation-based weight calculation method ensures that the closer the observation point is to the interpolation point, the better. The greater the contribution of the selection probability of an observation point to the interpolation point, the smoother and more reasonable probability transfer is achieved in space. Based on this, the selection probabilities of each observation point are further weighted and fused according to the spatial influence weight to obtain the interpolation result of the interpolation point. More importantly, the scheme in this embodiment also generates an uncertainty estimate of the interpolation point by calculation, which not only provides the predicted probability value, but also quantifies the reliability of the predicted value, so that users can fully consider the confidence of the prediction result when making subsequent decisions. By integrating the interpolation results of all grid cells and their corresponding uncertainty estimates, a continuous probability distribution map with probability density and uncertainty information is generated.

[0089] Through the above technical solution, this application can adaptively determine the bandwidth parameter of the Gaussian kernel function based on the spatial distribution characteristics of points in the predicted coordinate matrix, thereby enabling the spatial interpolation process to better adapt to the local characteristics of the data and avoid interpolation deviations that may be caused by fixed bandwidth. At the same time, by taking the center point of each grid cell as the interpolation point and the coordinate point in the predicted coordinate matrix as the observation point, the spatial influence weight is calculated based on the Gaussian kernel function to achieve effective weighted fusion of the selection probability of discrete predicted points, generating a smoother and more accurate interpolation result. More importantly, while calculating the interpolation result, the solution of this embodiment also provides its uncertainty estimation, so that the generated continuous probability distribution map not only contains probability density information, but also associates the reliability information of the prediction, which has the effect of improving the transparency and practicality of the prediction result. This allows users to assess the confidence of the prediction while understanding the distribution of residents' travel demand, which can effectively reduce the shortcomings of traditional spatial interpolation methods in parameter selection and result reliability assessment.

[0090] Further, in one embodiment, step S323 includes: S3231: Set the selection probability corresponding to the interpolation point to be randomized; In this embodiment, setting the selection probability corresponding to the interpolation point as a random variable means that when performing spatial interpolation, the selection probability of each interpolation point is not regarded as a fixed value, but rather modeled as a random quantity with inherent uncertainty. It can be understood that the true value of the selection probability of the interpolation point may fluctuate within a range and follow a certain probability distribution. For example, it can be modeled as a random variable that follows a specific probability distribution, and its distribution parameters can be estimated based on the selection probability of surrounding observation points, spatial distance, and other relevant characteristics. Alternatively, it can be regarded as a non-parametric random variable, and its uncertainty can be characterized by sampling surrounding observation data or constructing an empirical distribution.

[0091] S3232: Based on a preset probability generation model, the probability distribution of random variables is generated through Monte Carlo random simulation; In this embodiment, the preset probability generation model refers to a mathematical model or computational framework used to describe and generate the probability distribution of the aforementioned random variables. The probability generation model can be parameterized; for example, if the probability is chosen to be modeled as a Gaussian random variable, the probability generation model can be a Gaussian distribution model, describing its distribution characteristics by estimating its mean and standard deviation. Alternatively, the probability generation model can also be a data-driven non-parametric model, such as kernel density estimation or empirical distribution functions, used to infer the underlying probability distribution from finite sample data without pre-setting a specific distribution form. Monte Carlo random simulation refers to a computational method that estimates numerical results through repeated random sampling, particularly suitable for solving problems that are difficult to solve analytically. In this step, a large number of random samples can be drawn from the preset probability generation model through Monte Carlo random simulation. Approximating the true probability distribution of a random variable, for example, for a parameterized probability distribution model, a large number of random numbers that conform to the probability distribution model can be generated, and the random numbers represent multiple possible values ​​of the probability of selecting the interpolation point. In more complex scenarios, the Markov chain Monte Carlo method can be combined to sample from the posterior distribution to handle more complex probability distribution models, thereby more accurately capturing the distribution characteristics of the random variable. Generating the probability distribution of a random variable refers to the distribution obtained through Monte Carlo simulation that can describe all possible values ​​of the probability of selecting the interpolation point and their corresponding probabilities. Here, the probability distribution can be a discrete probability histogram, constructed by statistically analyzing the frequency of each value in the Monte Carlo simulation; furthermore, it can also be an approximation of a continuous probability density function or cumulative distribution function, obtained by fitting or smoothing a large number of simulated samples.

[0092] S3233: Extract the mean from the probability distribution as the interpolation result of the interpolation point to be interpolated, and calculate its variance or confidence interval as an uncertainty estimate.

[0093] In this embodiment, extracting the mean as the interpolation result for the interpolation point means using the expected value of the probability distribution obtained through Monte Carlo simulation as the final predicted value of the interpolation point. This is usually achieved by calculating the arithmetic mean of all Monte Carlo simulation samples, where the mean represents the most likely or expected value of the random variable. Calculating its variance or confidence interval as an uncertainty estimate is an indicator that quantifies the reliability or fluctuation range of the interpolation result. The variance is obtained by calculating the variance of the Monte Carlo simulation samples, which reflects the dispersion of the data points relative to the mean. The larger the variance, the higher the uncertainty. The confidence interval is determined by sorting the simulation samples to establish a certain percentage interval. This interval contains the true value of the random variable with a specific probability. For example, the 2.5% and 97.5% quantiles can be calculated to obtain a 95% confidence interval, thus intuitively representing the fluctuation range of the prediction result.

[0094] Specifically, the solution in this application models the selection probability of the interpolation point as a random variable and uses a pre-defined probability generation model to perform Monte Carlo random simulation, thereby generating the probability distribution of this random variable. This approach breaks through the limitation of traditional interpolation, which only provides a single deterministic result, and instead provides a probability distribution that comprehensively describes the multiple possibilities of the selection probability of the interpolation point. Based on this, the mean is extracted from the generated probability distribution as the interpolation result, providing a robust central tendency estimate. By calculating the variance or confidence interval of the probability distribution, the uncertainty of the interpolation result can be quantified. This not only enables the prediction of areas where residents' travel demand may be concentrated, but also allows for the assessment of the reliability of the prediction. For example, in areas with sparse or highly volatile data, the uncertainty estimate may be high, prompting users to interpret the prediction results for that area more cautiously. Conversely, in areas with sufficient and stable data, the uncertainty estimate may be low.

[0095] Through the above technical solution, this application can provide prediction results and uncertainty estimates for each interpolation point in a continuous probability distribution map, enabling users to clearly understand the reliability of prediction results for each region when viewing the probability distribution heatmap. For example, in addition to the probability density represented by color depth, uncertainty can be represented by additional visual elements on the probability distribution heatmap, such as transparency, texture, or independent layers, thereby helping users identify which regions have more robust prediction results and which regions have higher uncertainty. Through this quantification and presentation of uncertainty, the transparency and credibility of prediction conclusion reports can be enhanced, avoiding misjudgments caused by possible biases in a single prediction value.

[0096] In one embodiment, the semantic analysis model includes a feature fusion layer, a pattern recognition layer, and a report generation layer. Step S60 includes: S61: The feature fusion layer constructs multimodal feature vectors for each first point of interest based on the selection probability matrix, the information of the first point of interest, and the associated business distribution data; In this embodiment, the semantic analysis model aims to interpret and reason about complex unstructured or semi-structured data to extract meaningful information. In this embodiment, the semantic analysis model includes a feature fusion layer, a pattern recognition layer, and a report generation layer. The feature fusion layer is a module layer in the semantic analysis model used to integrate data from different sources and modalities to form a unified feature representation. For example, the feature fusion layer can encode and concatenate numerical data, spatial coordinate data, and categorical data to generate multimodal feature vectors that comprehensively describe each first point of interest. The encoding and concatenation can be performed using vector concatenation or nonlinear fusion through complex neural network structures to capture potential correlations between different modalities.

[0097] The feature fusion layer selects probability values ​​from the probability matrix, spatial attributes from the first point of interest information, and category labels from the business distribution data for feature encoding and concatenation to construct multimodal feature vectors for each first point of interest.

[0098] Specifically, the feature fusion layer constructs multimodal feature vectors for each primary point of interest (POI), converting different types of data into a unified numerical representation for subsequent processing. Probability values ​​selected from the probability matrix can be directly used as numerical features or normalized. Spatial attributes in the POI information, such as latitude and longitude coordinates, can be encoded as two-dimensional vectors or represented by calculated distances, directions, and other derived features from the origin information. Category labels in the business distribution data, such as "catering," "shopping," and "entertainment," can be converted into binary vectors using one-hot encoding or mapped to a low-dimensional dense vector space using word embedding techniques to capture semantic relationships between categories. These encoded features are then concatenated to form a comprehensive multimodal feature vector, fully representing the multifaceted attributes of each POI.

[0099] S62: The pattern recognition layer analyzes the multimodal feature vectors based on a predefined self-attention mechanism, calculates the semantic correlation between the information of each first interest point, and identifies the set of dominant interest point categories; In this embodiment, the pattern recognition layer refers to the module layer in the semantic analysis model, which is used to identify patterns, associations and regularities in the data by analyzing multimodal feature vectors. For example, the pattern recognition layer can use clustering algorithms to group first points of interest with similar features into one category, or use classification algorithms to predict the attractiveness of first points of interest to residents' travel.

[0100] Specifically, the pattern recognition layer analyzes multimodal feature vectors based on a predefined self-attention mechanism, calculates the semantic correlation between each primary point of interest (POI), and identifies the dominant POI category set. The self-attention mechanism is a technique that calculates the interdependencies between different elements in the input sequence, allowing the model to consider the importance of all other elements in the sequence when processing each element. In this step, the self-attention mechanism can be used to evaluate the mutual influence and correlation between various POIs within the target area in attracting residents' travel. By calculating the similarity between the query vector and the key vector, the semantic correlation between each POI can be obtained. For example, a restaurant and a cinema may have a high semantic correlation because they often together constitute residents' entertainment consumption scenarios. Based on the calculated semantic correlation, the pattern recognition layer can identify the POI categories that have the greatest impact on residents' travel needs or form spatial clustering effects, i.e., the dominant POI category set. Furthermore, this process can be implemented by analyzing attention weights or combining them with clustering algorithms.

[0101] S63: The report generation layer generates a prediction conclusion report based on the set of dominant interest categories and their corresponding statistical features, using pre-set natural language templates for content filling and logical assembly.

[0102] In this embodiment, the report generation layer refers to the output module layer in the semantic analysis model, which is used to present the key information identified by the pattern recognition layer in a human-readable natural language form. It usually involves using a predefined template, filling the extracted data into the corresponding positions of the predefined template, and assembling it according to a certain logical structure to finally form a well-structured prediction conclusion report.

[0103] Specifically, the report generation layer, based on the dominant point of interest (POI) category set and its corresponding statistical features, uses pre-set natural language templates to fill in content and assemble logic to generate a prediction report. The dominant POI category set refers to the types of POIs that significantly influence residents' travel needs, identified by the pattern recognition layer, such as "commercial retail" and "catering and entertainment." Its corresponding statistical features can include the number, density, average selection probability, and spatial distribution concentration of the dominant POIs within the target area. The pre-set natural language templates refer to pre-designed report structures containing placeholders for filling in specific statistical data and analysis results. For example, a natural language template might include "Residents' travel needs in the target area are mainly concentrated in the [dominant POI category set] area, characterized by [statistical feature 1], [statistical feature 2], [statistical feature 3], etc." The report generation layer fills the dominant POI category set and statistical features output by the pattern recognition layer into these natural language templates according to pre-set logical rules, and performs appropriate language organization and polishing to ultimately generate a complete prediction report.

[0104] Specifically, the solution in this application constructs a multi-layered architecture for the semantic analysis model, including a feature fusion layer, a pattern recognition layer, and a report generation layer, to achieve in-depth interpretation of residents' travel demand prediction results. Specifically, the feature fusion layer first integrates heterogeneous information such as the selection probability matrix, primary interest point information, and business distribution data to provide a data foundation for subsequent semantic analysis. On this basis, the pattern recognition layer uses a self-attention mechanism to deeply explore the semantic relationships between primary interest points in the target area and identify the dominant interest point category set that truly drives residents' travel demand. Finally, the report generation layer structures and texts the received analysis results using a pre-set natural language template to generate a specific and interpretable prediction conclusion report.

[0105] Furthermore, as an example, the training process of a semantic analysis model is as follows: Training data construction: This can be achieved by collecting no fewer than 5,000 samples of travel demand analysis reports written by experts in the field of urban planning. Each report corresponds to a set of input data, including the selection probability matrix of a specific area, the list of points of interest in that area, and statistical data on the distribution of their business types. The expert reports are then annotated in a structured manner to extract key information elements, including dominant point of interest categories, statistical feature descriptions such as quantity, density, spatial distribution patterns, and causal analysis conclusions, to form labeled data for supervised learning.

[0106] Feature fusion layer training: The feature fusion layer can adopt a multi-layer perceptron structure, which includes three fully connected layers with hidden layer dimensions of 256, 128 and 64 respectively. The activation function is ReLU. Its input is the concatenated multimodal feature vector, which includes the encoded vector of probability value, spatial coordinates and class label. The output is a 64-dimensional fused feature vector.

[0107] Pattern recognition layer training: The pattern recognition layer can adopt the Transformer encoder structure, which contains 4 layers of self-attention modules, with 8 attention heads in each layer and a feedforward network dimension of 512. Through the self-attention mechanism, the model can automatically learn the semantic association strength between each point of interest and identify the set of interest point categories that have a dominant influence on residents' travel needs.

[0108] Report generation layer training: The report generation layer can adopt a sequence-to-sequence architecture, and the decoder is an autoregressive language model based on the GPT structure, which contains 12 Transformer decoders and has a hidden layer dimension of 768.

[0109] Loss function design: The loss function can be the standard cross-entropy loss of the language model, which includes conditional probability, and the conditional probability can be calculated by the decoder.

[0110] Training strategy: The AdamW optimizer can be used, with a learning rate set to 5×10. -5 The batch size is 32, the training rounds are 30, and the BLEU score and ROUGE score are monitored on the validation set. Training stops when both metrics no longer improve.

[0111] It should be noted that those skilled in the art can reproduce and train the semantic analysis model based on the examples and content described in this application specification, combined with actual needs.

[0112] Through the above technical solutions, this application can transform abstract prediction probabilities into concrete and interpretable semantic information, effectively solving the problems of generalized and shallow prediction conclusion reports. In particular, through the hierarchical architecture of feature fusion, pattern recognition, and report generation, the semantic analysis model can deeply explore the driving factors behind residents' travel needs, identify the dominant interest categories that have a key impact on travel choices, and clearly present them in the form of natural language. This provides more easily understood and adopted decision support for planners or operators, and has the effect of improving the practicality and value of prediction results.

[0113] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0114] In one embodiment, a resident travel demand prediction system based on user profiles is provided. This resident travel demand prediction system based on user profiles corresponds one-to-one with the resident travel demand prediction method based on user profiles described in the previous embodiment. The resident travel demand prediction system based on user profiles includes: The constraint acquisition module is used to respond to the demand prediction command input by the user terminal and the corresponding travel origin information, determine the target area information and target group, and acquire the prediction constraints associated with the demand prediction command. The prediction constraints include travel time information, travel mode information and travel population profile tags. The matrix generation module is used to input the prediction constraints into the pre-trained prediction model and generate the prediction coordinate matrix and its corresponding selection probability matrix. The heatmap generation module is used to generate a continuous probability distribution map associated with target area information based on the predicted coordinate matrix and the selection probability matrix through a predefined spatial interpolation strategy, and render the continuous probability distribution map as a probability distribution heatmap in an interactive scene. The boundary determination module is used to determine the geometric boundary information of the target area in response to the user's interactive selection operation on the probability distribution heatmap. The data acquisition module is used to identify all first points of interest within the target area based on geometric boundary information, and to acquire business distribution data associated with each first point of interest. The report generation module is used to input the selection probability matrix, the first point of interest information and its associated business distribution data into a pre-trained semantic analysis model to generate a prediction conclusion report.

[0115] For specific limitations regarding a user profile-based resident travel demand prediction system, please refer to the limitations of a user profile-based resident travel demand prediction method described above, which will not be repeated here. Each module in the aforementioned user profile-based resident travel demand prediction system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0116] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for predicting residents' travel demand based on user profiles, characterized in that, Including the following steps: In response to the demand forecasting command input by the user and the corresponding travel origin information, the target area information and target group are determined, and the forecasting constraints associated with the demand forecasting command are obtained. The forecasting constraints include travel time information, travel mode information and travel population profile tags. Input the prediction constraints into the pre-trained prediction model to generate the prediction coordinate matrix and its corresponding selection probability matrix; Based on the predicted coordinate matrix and the selection probability matrix, a continuous probability distribution map associated with the target area information is generated through a predefined spatial interpolation strategy, and the continuous probability distribution map is rendered as a probability distribution heatmap in an interactive scene. In response to the user's interactive selection operation of the target area in the probability distribution heatmap, determine the geometric boundary information of the target area; Based on geometric boundary information, identify all first points of interest within the target area and obtain business distribution data associated with each first point of interest; The selected probability matrix, the first point of interest information, and the associated business distribution data are input into a pre-trained semantic analysis model to generate a prediction conclusion report.

2. The method for predicting residents' travel demand based on user profiles according to claim 1, characterized in that: The prediction model The model includes a feature embedding layer, a spatial attention layer, and a probability output layer. The step of inputting prediction constraints into a pre-trained prediction model to generate a prediction coordinate matrix and its corresponding selection probability matrix includes the following steps: The feature embedding layer maps the received prediction constraints into corresponding dense feature vectors, and concatenates and nonlinearly fuses the dense feature vectors to generate scene feature vectors. The spatial attention layer uses the origin information of the trip as the query center and the information of the second points of interest within a preset range around the query center as the key value to construct a spatial attention graph. It also generates the spatial association weight of each second point of interest relative to the origin information by calculating the semantic association degree between the query center and the key value and the geographical decay factor. The probability output layer sorts and filters the second interest points based on spatial correlation weights, and performs probability normalization on the spatial correlation weights of the filtered second interest point set, outputting the predicted coordinate matrix and its corresponding selection probability matrix.

3. The method for predicting residents' travel demand based on user profiles according to claim 2, characterized in that: The feature embedding layer maps the received prediction constraints to corresponding dense feature vectors, and performs concatenation and nonlinear fusion of the dense feature vectors to generate scene feature vectors, including the following steps: Travel time information is mapped into periodic temporal embedding vectors, and traveler profile tags are mapped into semantic embedding vectors. Based on travel mode information and origin information, calculate the distance decay curve that can be reached within a preset time period and encode it as a spatial reachability embedding vector. The temporal periodicity embedding vector, semantic embedding vector, and spatial reachability embedding vector are concatenated to form a scene feature vector.

4. The method for predicting residents' travel demand based on user profiles according to claim 2, characterized in that: The spatial attention layer constructs a spatial attention graph using the origin information as the query center and the information of second points of interest within a preset range around the query center as the key. It then generates the spatial association weights of each second point of interest relative to the origin information by calculating the semantic correlation between the query center and the key, as well as the geographical attenuation factor. The steps include: Based on scene feature vectors, query vectors representing travel origin information, key vectors representing information of each second point of interest, and value vectors carrying semantic information are generated through independent linear transformation layers. The initial semantic association between each second point of interest and the origin information is obtained by calculating the cosine similarity between the query vector and each key vector. Obtain the spherical distance between the origin of travel and each second point of interest, and generate the geographic attenuation factor corresponding to the spherical distance through a preset attenuation function; The initial semantic relevance of each second point of interest is weighted by its corresponding geographic decay factor, and all second points of interest are fused and normalized to generate the spatial relevance weight of each second point of interest relative to the travel origin information.

5. The method for predicting residents' travel demand based on user profiles according to claim 1, characterized in that: The step of generating a continuous probability distribution map associated with target region information based on the predicted coordinate matrix and the selection probability matrix using a predefined spatial interpolation strategy, and rendering the continuous probability distribution map as a probability distribution heatmap in an interactive scene, includes the following steps: The grid division granularity is determined based on the spatial distribution density of points in the predicted coordinate matrix by using a pre-set dynamic granularity determination strategy, and several grid cells are generated according to the grid division granularity. The Gaussian kernel interpolation algorithm is used to perform spatial interpolation on the predicted coordinate matrix and the selection probability matrix on the generated grid cells to generate a continuous probability distribution map. By associating a continuous probability distribution map with the real-time viewpoint height of the interactive scene and dynamically adjusting the rendering granularity and color transparency, a probability distribution heatmap is generated.

6. The method for predicting residents' travel demand based on user profiles according to claim 5, characterized in that: The step of determining the mesh granularity based on the spatial distribution density of points in the predicted coordinate matrix using a pre-set dynamic granularity determination strategy, and generating several mesh elements according to the mesh granularity, includes the following steps: Calculate the spatial distribution density of all coordinate points in the predicted coordinate matrix within the target area, and identify high-density clustered areas and low-density sparse areas. A fine-grained grid partitioning strategy is adopted for high-density clustered areas, and a coarse-grained grid partitioning strategy is adopted for low-density sparse areas to determine a multi-scale grid partitioning scheme. Based on the multi-scale meshing scheme, an adaptive mesh including mesh units of different granularities is generated.

7. The method for predicting residents' travel demand based on user profiles according to claim 5, characterized in that: The step of generating a continuous probability distribution map by performing spatial interpolation processing on the predicted coordinate matrix and the selection probability matrix on the generated grid cells using the Gaussian kernel function interpolation algorithm includes the following steps: Based on the spatial distribution characteristics of points in the predicted coordinate matrix, the bandwidth parameter of the Gaussian kernel function is determined. The center point of each grid cell is taken as the interpolation point, and the coordinate point in the predicted coordinate matrix is ​​taken as the observation point. The spatial influence weight of each observation point on the interpolation point is calculated based on the Gaussian kernel function. Based on the spatial influence weight, the selection probabilities corresponding to each observation point are weighted and fused, and the interpolation results and uncertainty estimates of the interpolation points to be interpolated are calculated and generated. By integrating the interpolation results and uncertainty estimates of each grid cell, a continuous probability distribution map with associated probability density and uncertainty information is generated.

8. The method for predicting residents' travel demand based on user profiles according to claim 7, characterized in that: The step of weighting and fusing the selection probabilities corresponding to each observation point based on spatial influence weights, and calculating and generating the interpolation result and its uncertainty estimate for the interpolation point to be interpolated, includes the following steps: Set the selection probability corresponding to the interpolation point to be randomized; Based on a pre-defined probability generation model, the probability distribution of random variables is generated through Monte Carlo random simulation. Extract the mean from the probability distribution as the interpolation result of the interpolation point to be interpolated, and calculate its variance or confidence interval as an uncertainty estimate.

9. The method for predicting residents' travel demand based on user profiles according to claim 1, characterized in that: The semantic analysis model includes a feature fusion layer, a pattern recognition layer, and a report generation layer. The step of inputting the selection probability matrix, the first point of interest information, and its associated business distribution data into the pre-trained semantic analysis model to generate a prediction conclusion report includes the following steps: The feature fusion layer constructs multimodal feature vectors for each first point of interest based on the selection probability matrix, the information of the first point of interest, and the associated business distribution data; The pattern recognition layer analyzes the multimodal feature vectors based on a predefined self-attention mechanism, calculates the semantic correlation between the information of each first interest point, and identifies the set of dominant interest point categories. The report generation layer generates a predictive conclusion report based on the set of dominant interest categories and their corresponding statistical features, using pre-set natural language templates for content filling and logical assembly.

10. A resident travel demand prediction system based on user profiles, characterized in that, include: The constraint acquisition module is used to respond to the demand prediction command input by the user terminal and the corresponding travel origin information, determine the target area information and target group, and acquire the prediction constraints associated with the demand prediction command. The prediction constraints include travel time information, travel mode information and travel population profile tags. The matrix generation module is used to input the prediction constraints into the pre-trained prediction model and generate the prediction coordinate matrix and its corresponding selection probability matrix. The heatmap generation module is used to generate a continuous probability distribution map associated with target area information based on the predicted coordinate matrix and the selection probability matrix through a predefined spatial interpolation strategy, and render the continuous probability distribution map as a probability distribution heatmap in an interactive scene. The boundary determination module is used to determine the geometric boundary information of the target area in response to the user's interactive selection operation on the probability distribution heatmap; The data acquisition module is used to identify all first points of interest within the target area based on geometric boundary information, and to acquire business distribution data associated with each first point of interest. The report generation module is used to input the selection probability matrix, the first point of interest information and its associated business distribution data into a pre-trained semantic analysis model to generate a prediction conclusion report.

Images