Lightweight transportation junction passenger flow prediction system and method

Abstract

The invention discloses a lightweight transportation junction passenger flow prediction system and method, and relates to the technical field of artificial intelligence and traffic management. The lightweight transportation junction passenger flow prediction system comprises a multi-source data acquisition module, a data processing and enhancing module and a lightweight CNN model construction module. The multi-source data acquisition module is used for acquiring historical passenger flow data, population migration data and other auxiliary data; the data processing and enhancing module is used for outputting a training set, a verification set and a test set; the lightweight CNN model construction module is used for realizing passenger flow prediction. The lightweight transportation junction passenger flow prediction system method comprises the following steps: collecting multi-source data; preprocessing and enhancing data; constructing a lightweight CNN model; training and optimizing the model; and obtaining a prediction result. The method has the advantages that the lightweight passenger flow prediction model is constructed, the defects of an existing method in the aspects of long-time prediction, calculation efficiency and inter-city interactive modeling are overcome, and high-precision and low-cost traffic hub passenger flow prediction is achieved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and traffic management technology, and in particular to a lightweight transportation hub passenger flow prediction system and method. Background Art

[0002] With the acceleration of urbanization and the expansion of transportation networks, passenger flow forecasting at transportation hubs (such as railway stations and airports) has become a key technology for optimizing resource allocation and improving operational efficiency.

[0003] Traditional passenger flow prediction methods mainly include the following two categories: 1. Traditional empirical forecasting: It relies on manual statistics and subjective experience and lacks quantitative analysis of multi-source data (such as weather, holidays, and inter-city interactions). This leads to low forecast accuracy, especially in long-term (such as monthly and quarterly) and emergency event scenarios (such as typhoons and military disasters).

[0004] 2. Deep learning time series prediction models: Such as long short-term memory networks (LSTMs) and Transformers. Although they can capture time series dependencies, they suffer from high computational complexity (large number of model parameters, high training and inference time costs, and difficulty meeting the real-time prediction needs of transportation hubs), neglect of inter-city interaction effects (without integrating population migration data, it is impossible to model the dynamic impact of inter-city population mobility on transportation hub passenger flow), and insufficient frequency domain feature mining (based on time domain modeling, it is difficult to effectively extract the global periodic characteristics of passenger flow data, such as holiday patterns and seasonal fluctuations).

[0005] In summary, traditional passenger flow forecasting methods have problems such as insufficient prediction accuracy (traditional methods and deep learning models are easily affected by noise in long-term predictions, resulting in inaccuracy), high model complexity (complex models have high hardware resource requirements and are difficult to deploy on low-cost equipment), and limited model interpretability and predictive ability (external data such as population migration are not integrated, and the model interpretability is limited). Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a lightweight transportation hub passenger flow prediction system and method, which solves the problems of traditional passenger flow prediction methods such as insufficient prediction accuracy, high model complexity, and limited model interpretability.

[0007] The technical solution of the present invention is: a lightweight transportation hub passenger flow prediction system, comprising a multi-source data acquisition module, a data processing and enhancement module and a lightweight CNN model construction module; The multi-source data acquisition module is used to collect historical passenger flow data, population migration data and other auxiliary data; The data processing and enhancement module processes the raw data collected by the multivariate data acquisition module and outputs training, validation, and test sets. The data processing and enhancement module includes a data cleaning module, a feature engineering module, a feature screening module, a data enhancement module, and a data post-processing module. The data enhancement module combines population migration data with a preset population migration ratio threshold to screen cities with strong correlations with the target hub and construct a population in-migration / out-migration matrix based on this data. The lightweight CNN model construction module is used to realize passenger flow prediction; the lightweight CNN model construction module includes an adaptive wavelet analysis module, a CNN network module, a data fusion module, a time domain signal reconstruction module and a result output module; the data fusion module is used to weightedly fuse the population in / out matrix with the frequency domain component.

[0008] A further technical solution of the present invention is: in the data processing and enhancement module: a data cleaning module is used to process missing values ​​and outliers in the data; a feature engineering module is used to convert the data processed by the data cleaning module into numerical features suitable for model input; a feature screening module screens out multiple features that have the greatest impact on passenger flow based on the numerical features converted by the feature engineering module; a data post-processing module is used to normalize the data processed by the data cleaning module, the feature engineering module, the feature screening module, and the data enhancement module, and then divide the normalized data into a training set, a validation set, and a test set according to time windows, and finally convert the data into a model input format; In the lightweight CNN model construction module: the CNN network module is trained on the normalized data output by the data post-processing module; the network structure of the CNN network module includes a one-dimensional convolution layer; the one-dimensional convolution layer uses depthwise separable convolution to decompose the standard convolution into depthwise convolution and point-by-point convolution, reducing the number of parameters to 1 / N of the traditional convolution, where N is the number of channels; the time domain signal reconstruction module is used to map the weighted fused frequency domain signal back to the time domain through an inverse wavelet transform to form the final passenger flow prediction value; the result output module is used to output the passenger flow prediction result including the date, predicted passenger flow and confidence interval.

[0009] A further technical solution of the present invention is that other auxiliary data also include temperature data, air quality data and health-related data.

[0010] The technical solution of the present invention is: a lightweight transportation hub passenger flow prediction method, based on a lightweight transportation hub passenger flow prediction system, the method is as follows: S01, multi-source data acquisition: The following data is collected through the multi-source data collection module: historical passenger flow data: including the number of people arriving and departing from the station, train number and timestamp; population migration data: the percentage of population inflow / outflow between the transportation hub and other cities outside the transportation hub city; other auxiliary data: including but not limited to weather data and holiday data; S02, data preprocessing and enhancement: A. Data cleaning: The data cleaning module processes missing values ​​and outliers in the collected multi-source data, fills in missing values, and removes outliers to ensure data integrity and reliability. B. Feature Engineering: The feature engineering module converts cleaned data into numerical features suitable for model input; one-hot encoding is used for unordered categorical variables; label encoding is used for ordered categorical variables; C. Feature screening: The feature screening module filters the numerical features converted by the feature engineering module to select the features that have the greatest impact on customer flow; D. Data Enhancement: Import population migration data into the data enhancement module. Based on the preset population migration ratio threshold, select cities with strong correlation with the target hub and construct a population in-migration / out-migration matrix. S03, lightweight CNN model construction: A. Frequency Domain Feature Extraction: Adaptive wavelet analysis module is used to implement the following operations: Wavelet decomposition: Adaptive wavelet function is used to decompose time domain passenger flow data into frequency domain components, extracting high-frequency and low-frequency features respectively; Frequency Domain Feature Meshing: Two independent one-dimensional convolutional neural networks are preset to fit the extracted high-frequency and low-frequency components respectively; Nonlinear Enhancement: The GELU operator is used to enhance the nonlinear expression capability of features; B. CNN model construction: The constructed CNN network module includes a one-dimensional convolutional layer and a parameter sharing strategy. The one-dimensional convolutional layer uses depthwise separable convolution to decompose the standard convolution into depthwise convolution and pointwise convolution. The parameter sharing strategy shares the convolution kernel weights across multiple time windows. C. Data fusion: The data fusion module is used to fuse the population in-migration / out-migration matrix with the frequency domain components in a weighted manner to quantify the contribution of population mobility to passenger flow fluctuations; S04, model training and optimization: A. Model training: The backbone network is pre-trained based on historical passenger flow data to enable the model to initially learn the basic temporal patterns of passenger flow. The loss function uses mean square error and the Adam optimizer is used. B. Model Optimization: The population in-migration / out-migration matrix and the multiple features with the greatest impact on passenger flow identified in step S02 are used as augmented data to retrain the pre-trained backbone network to optimize the model's ability to model inter-city interactions and external events. S05, get the prediction result: A. Time Domain Signal Reconstruction: After weighted fusion of the population in-migration / out-migration matrix and the frequency domain components, a time domain prediction sequence is generated through inverse wavelet transform, and a sequence of passenger flow prediction values ​​for future time periods is output; B. Forecast result output: The result output module outputs the forecast results in JSON or CSV format, including the date, predicted passenger flow, and confidence interval.

[0011] A further technical solution of the present invention is: in step S01, weather data includes weather conditions and extreme weather events, which are sourced from the Meteorological Bureau or a third-party weather platform; temperature data includes daily average temperature and daily extreme temperature and corresponding time, and the data source is the Meteorological Bureau or a third-party weather platform; holiday data includes the date and length of statutory holidays, as well as the date and length of special holidays, and the data source is the government public calendar and education department announcements.

[0012] A further technical solution of the present invention is: in step S01, other auxiliary data also include temperature data, air quality data and health-related data; temperature data includes daily average temperature, daily temperature extremes and corresponding time, and the data source is the Meteorological Bureau or a third-party weather platform; air quality data includes pollutant concentration values ​​and air quality levels, and the data source is the environmental protection department monitoring station or a third-party environmental data platform; health-related data includes regional health risk levels and travel restrictions, and the data source is the National Health Commission announcement and the local government health prevention and control platform.

[0013] A further technical solution of the present invention is: in step S02, the in-migration / out-migration matrix includes an in-migration matrix and an out-migration matrix; the in-migration matrix represents the percentage of population in-migration from each city to the target transportation hub, and the out-migration matrix represents the percentage of population out-migration from the target transportation hub to each city; in step S02, the migration ratio threshold includes an in-migration ratio threshold and an out-migration ratio threshold, and satisfying any one of these two thresholds is deemed to satisfy the migration ratio threshold; the in-migration ratio threshold: the average daily population in-migration from the city to the target hub accounts for ≥1% of the total in-migration of the target hub; the out-migration ratio threshold: the average daily population out-migration from the target hub to the city accounts for ≥1% of the total out-migration of the target hub.

[0014] A further technical solution of the present invention is: in step S02, the feature screening operation is as follows: first, Spearman correlation analysis is used to calculate the correlation coefficient between each feature and passenger flow, and based on the rule that the absolute value of the Spearman correlation coefficient satisfies |ρ|≥0.5, strongly correlated features are retained, and based on the rule that the absolute value of the Spearman correlation coefficient satisfies |ρ|<0.3, weakly correlated features are eliminated; then, a recursive feature elimination method is used, based on the random forest model, to evaluate the importance of features by decreasing the Gini index, and the feature with the lowest importance ranking is eliminated in each iteration, and the mean square error of the random forest model is monitored by 5-fold cross validation. When the mean square error drops by more than 5%, the elimination is terminated, and the top 10 features in importance are retained.

[0015] A further technical solution of the present invention is: in step S02, the scheme for filling missing values ​​is as follows: for numerical features, if the missing ratio is less than 5%, the mean is used for filling; if the missing ratio is ≥5%, 0 is used for filling or interpolation of previous and next values; for categorical features, the mode is used for filling; for missing periods in time series features, linear interpolation or neighboring city data is used to replace them; in step S02, the scheme for eliminating outliers is as follows: outliers are eliminated by combining statistical methods and business rules; for numerical features, the 3σ principle is used to eliminate outliers, and then unreasonable data is eliminated by combining domain knowledge.

[0016] A further technical solution of the present invention is: in step S03, the network structure of the CNN network module includes, in addition to the one-dimensional convolutional layer, an input layer, a pooling layer and a fully connected layer; the input layer is used to receive frequency domain components and the population in / out matrix; the pooling layer uses maximum pooling to compress the feature dimension and retain significant features; the number of output neurons in the fully connected layer is consistent with the prediction time span.

[0017] A further technical solution of the present invention is: in step S04, during the second training process, the convolutional layer weights of the low-frequency components in the backbone network are fixed, and only the weights of the high-frequency components and the data enhancement channel are updated; in step S04, during the two training processes, the gradient of the loss function with respect to the model parameters is calculated by automatic differentiation, and the model weight parameters are iteratively updated based on the gradient direction.

[0018] Compared with the prior art, the present invention has the following advantages: 1. It constructs a lightweight passenger flow prediction model through "lightweight CNN + wavelet frequency domain analysis + population migration data enhancement", which solves the shortcomings of existing methods in long-term prediction, computational efficiency, and inter-city interaction modeling, and achieves high-precision and low-cost passenger flow prediction for transportation hubs.

[0019] 2. High-Precision Prediction Capabilities: Using adaptive wavelet analysis, the model decomposes time-domain passenger flow data into frequency-domain components, effectively capturing long-term cyclical characteristics (such as holiday patterns and seasonal fluctuations) while reducing noise interference. By incorporating intercity migration data, the model quantifies the dynamic impact of intercity population flows on transportation hubs, enhancing its ability to model complex interactions. Experimental data shows that incorporating intercity migration data reduces the mean squared error (MSE) of predictions to 0.122 and 0.356 for 32-day and 64-day timeframes, respectively, significantly outperforming traditional LSTM and Transformer models.

[0020] 3. Lightweight model and efficient computation: I. Lightweight CNN design: Using one-dimensional convolutional layers and parameter sharing strategies to reduce the number of model parameters, while combining adaptive wavelet analysis to replace traditional redundant convolution operations. II. Frequency-time domain decoupling: After completing the main feature modeling in the frequency domain, only the inverse wavelet transform is required to reconstruct the time domain prediction results, greatly reducing the computational complexity. Based on the above two technical means, the model training time cost is reduced by approximately 40%, the inference speed is increased by more than 2 times, and it can be deployed on low-cost hardware (such as edge computing devices).

[0021] 4. Multi-source data fusion for enhanced interpretability: This model integrates historical passenger flow data, population migration data, weather, holidays, and other multi-source information to construct global feature inputs. Furthermore, through Spearman coefficient analysis, it screens for strongly correlated features (such as train numbers and immigration and emigration percentages) to avoid redundant data interference. These two technical approaches enhance the model's interpretability, for example, by clearly quantifying the negative impact of typhoons on passenger flow (with a correlation coefficient of -0.0636), providing transparent data for operational decision-making at transportation hubs.

[0022] 5. Real-time and Scalability: Supports dynamic access to real-time data (such as sudden weather events) to achieve near-real-time forecasting. The modular design allows for rapid expansion to other transportation hubs (such as airports and long-distance bus stations) by simply adjusting the data interface.

[0023] 6. Resource Optimization and Cost Reduction: Based on a lightweight model, this system reduces GPU memory requirements and supports deployment on standard servers or the cloud. The model also includes a built-in performance evaluation module that dynamically adjusts hyperparameters, reducing manual intervention costs. This technically reduces deployment costs by over 50%. Furthermore, through accurate predictions, it optimizes resource allocation (such as manpower scheduling and energy allocation) at transportation hubs, reducing operational waste.

[0024] The present invention is further described below with reference to the figures and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a structural schematic diagram of the present invention; Figure 2 Flow chart of the method of the present invention.

[0026] Legend: Multi-source data acquisition module 1; historical passenger flow data acquisition module 11; population migration data acquisition module 12; other auxiliary data acquisition module 13; data processing and enhancement module 2; data cleaning module 21; feature engineering module 22; feature screening module 23; data enhancement module 24; data post-processing module 25; lightweight CNN model construction module 3; adaptive wavelet analysis module 31; CNN network module 32; data fusion module 33; time domain signal reconstruction module 34; result output module 35. DETAILED DESCRIPTION

[0027] Example 1 Figure 1 As shown in Figure 1, the lightweight transportation hub passenger flow prediction system includes a multi-source data acquisition module 1, a data processing and enhancement module 2, and a lightweight CNN model construction module 3. The multi-source data acquisition module is used to collect historical passenger flow data from the transportation hub, as well as population migration data and other auxiliary data from external data sources. The data processing and enhancement module receives and processes the raw data collected by the multi-source data acquisition module and outputs standardized training sets, test sets, and validation sets. The lightweight CNN model construction module achieves (high-precision, low-complexity) passenger flow prediction through frequency domain modeling, module construction, and data fusion operations.

[0028] The multi-source data collection module 1 includes a historical passenger flow data collection module 11 , a population migration data collection module 12 and other auxiliary data collection modules 13 .

[0029] The historical passenger flow data collection module 11 is used to collect historical passenger flow data. The data source of the historical passenger flow data is the internal system of the transportation hub (such as Hangzhou East Station). The data type includes the number of train arrivals and departures, train number and timestamp.

[0030] The population migration data collection module 12 is used to collect population migration data. The data source of the population migration data is an external data source (such as Baidu's population migration big data platform), and the data type is the percentage of population inflow / outflow between the transportation hub and other cities (excluding other cities other than the city where the transportation hub is located).

[0031] The other auxiliary data collection module 13 is used to collect other auxiliary data, including weather data, temperature data, air quality data, holiday data, and health-related data. Weather data types include weather conditions (sunny, cloudy, rainy, snowy, etc.) and extreme weather events (typhoons, heavy rain, blizzards, etc.). The data source is the Meteorological Bureau or a third-party weather platform. Temperature data types include daily average temperature, daily maximum temperature and corresponding time, and daily minimum temperature and corresponding time. The data source is the Meteorological Bureau or a third-party weather platform. Air quality data types include concentrations of pollutants such as PM2.5, PM10, SO2, and NO2, and air quality levels (excellent, good, slightly polluted, etc.). The data source is environmental protection department monitoring stations or third-party environmental data platforms (such as AQICN). Holiday data types include the dates and lengths of statutory holidays (Spring Festival, Labor Day, National Day, etc.), as well as the dates and lengths of special holidays (local holidays, school winter and summer vacations, etc.). The data source is the government calendar and education department announcements. The data types of health-related data include regional health risk levels and travel restrictions. The data sources are announcements from the National Health Commission and local government health prevention and control platforms.

[0032] The data processing and enhancement module 2 includes a data cleaning module 21 , a feature engineering module 22 , a feature screening module 23 , a data enhancement module 24 and a data post-processing module 25 .

[0033] The data cleaning module 21 is used to process missing values ​​and outliers in the data to ensure data integrity and reliability. Missing value filling: For numerical features (such as temperature, air quality index, etc.), if the missing ratio is less than 5%, the mean filling is used (such as historical average temperature). If the missing ratio is ≥5%, 0 filling or interpolation of previous and next values ​​is used; for categorical features (such as weather, holidays, etc.), the mode filling is used; for missing periods in time series features (such as population migration data), linear interpolation or neighboring city data are used to replace them. Outlier removal: Statistical methods and business rules are combined to remove outliers; for numerical features (such as the number of train arrivals and departures, weather, etc.), the 3σ principle (values ​​exceeding the mean ±3 times the standard deviation are considered outliers) is used to remove outliers, and domain knowledge is combined to eliminate unreasonable data (such as a 99% drop in passenger flow during heavy rain).

[0034] The feature engineering module 22 converts the data processed by the data cleaning module into numerical features suitable for model input. Unordered categorical variables (such as weather types: sunny, rainy, overcast, and snowy) are encoded using one-hot encoding (for example, "sunny" is encoded as [1, 0, 0, 0], and "rainy" is encoded as [0, 1, 0, 0]). Ordered categorical variables are encoded using label encoding (such as days of the week: 1-7).

[0035] Based on the numerical features converted by the feature engineering module, the feature screening module 23 selects multiple features (variables) with the greatest impact on passenger flow (the target variable). First, a Spearman correlation analysis is performed to calculate the correlation coefficient between each feature and passenger flow. Strongly correlated features (such as train number, with a correlation coefficient of 0.803) are retained, based on the rule that the absolute value of the Spearman correlation coefficient satisfies |ρ| ≥ 0.5. Weakly correlated features (such as the air quality index, with a correlation coefficient of 0.027) are removed, based on the rule that the absolute value of the Spearman correlation coefficient satisfies |ρ| < 0.3. Then, a recursive feature elimination method is used, based on a random forest model, to assess feature importance by decreasing the Gini index. Each iteration, the lowest-ranked feature is eliminated. The mean squared error of the random forest model is monitored through 5-fold cross-validation. Elimination is terminated when the mean squared error decreases by more than 5%, retaining the top 10 most important features.

[0036] Data enhancement module 24 improves model generalization and prediction accuracy by introducing population migration data. This data is obtained from Baidu's population migration big data platform. A population in-migration / out-migration matrix is ​​then constructed based on this data. The in-migration matrix represents the percentage of population inflow from each city to the target transportation hub, while the out-migration matrix represents the percentage of population outflow from the target transportation hub to each city. Population migration data is updated daily to reflect real-time population mobility trends. Cities with a strong correlation to the target hub are screened based on preset population migration ratio thresholds (in-migration ratio: the average daily inflow from the city to the target hub as a percentage of the total inflow to the target hub is ≥1%; out-migration ratio: the average daily outflow from the target hub to the city as a percentage of the total outflow from the target hub is ≥1%). If population migration data is missing, cities with a straight-line distance to the target hub of ≤300 km and a DGP interaction strength of ≥5% are screened. The in-migration / out-migration matrix is ​​constructed based on these screened cities and used in subsequent model training and prediction processes. Finally, the in-migration / out-migration matrix is ​​aligned with the passenger flow data by timestamp to ensure temporal consistency.

[0037] The data post-processing module 25 is used to normalize the data to eliminate the dimensional differences between features, then divide the normalized data into training sets, validation sets and test sets according to time windows, and finally convert the data into model input format.

[0038] The lightweight CNN model construction module 3 includes an adaptive wavelet analysis module 31, a CNN network module 32, a data fusion module 33, a time domain signal reconstruction module 34 and a result output module 35.

[0039] The adaptive wavelet analysis module 31 is used to implement the following operations: 1. Wavelet decomposition: Use the adaptive wavelet function to decompose the time domain passenger flow data into frequency domain components, and extract high-frequency (detail fluctuations) and low-frequency (long-term trend) features; 2. Frequency domain feature meshing: Preset two independent one-dimensional convolutional neural networks to fit the extracted high-frequency and low-frequency components respectively; 3. Nonlinear enhancement: Enhance the nonlinear expression ability of features through GELU and or functions.

[0040] The network structure of CNN network module 32 consists of an input layer, a one-dimensional convolutional layer, a pooling layer, and a fully connected layer. The input layer receives frequency domain components and a population in-migration / out-migration matrix. The one-dimensional convolutional layer uses depthwise separable convolution, decomposing the standard convolution into depthwise convolution and pointwise convolution, reducing the number of parameters to 1 / N (N is the number of channels) of traditional convolution. The convolution kernel size is 3×1, with a stride of 1 and padding of 1. The CNN network module uses a parameter sharing strategy, sharing convolution kernel weights across multiple time series windows to avoid redundant parameters (for example, day of week features of different dates share the same convolution kernel). The pooling layer uses max pooling (pooling window 2×1) to compress feature dimensions and retain significant features. The number of output neurons in the fully connected layer corresponds to the prediction time span (for example, a 16-day prediction corresponds to 16 neurons).

[0041] The data fusion module 33 is used to perform a weighted fusion of the population in-migration / out-migration matrix and the frequency domain component, thereby quantifying the contribution of population mobility to passenger flow fluctuations (for example, for every 1% increase in in-migration from Shanghai, the passenger flow at the target hub increases by 0.6%).

[0042] The time domain signal reconstruction module 34 is used to map the weighted fused frequency domain signal back to the time domain through inverse wavelet transform to form the final passenger flow prediction value.

[0043] The result output module 35 is used to output the prediction results in JSON or CSV format, including the date, predicted passenger flow and confidence interval.

[0044] like Figure 2 As shown, a lightweight transportation hub passenger flow prediction method is based on the above-mentioned lightweight transportation hub passenger flow prediction system, and the method is as follows: S01, multi-source data acquisition: The following data is collected through the multi-source data collection module: I. Historical passenger flow data: including the number of people arriving and departing from the station, train numbers and timestamps; II. Population migration data: the percentage of population inflow / outflow between the transportation hub and other cities outside the city where the transportation hub is located; III. Other auxiliary data: including but not limited to weather data, temperature data and holiday data.

[0045] In this step, population migration data comes from Baidu's population migration big data platform.

[0046] In this step, weather data includes weather conditions and extreme weather events, and comes from the Meteorological Bureau or a third-party weather platform; temperature data includes daily average temperature and daily extreme temperature and corresponding time, and the data source is the Meteorological Bureau or a third-party weather platform; holiday data includes the date and length of statutory holidays, as well as the date and length of special holidays, and the data source is the government public calendar and education department announcements.

[0047] In this step, auxiliary data also include air quality data and health-related data; air quality data includes pollutant concentration values ​​and air quality levels, and the data source is the environmental protection department's monitoring station or a third-party environmental data platform; health-related data includes regional health risk levels and travel restrictions, and the data source is the National Health Commission's announcement and the local government's health prevention and control platform.

[0048] S02, data preprocessing and enhancement: A. Data cleaning: The data cleaning module processes missing values ​​and outliers in the collected multi-source data, fills in missing values, and removes outliers to ensure data integrity and reliability. B. Feature Engineering: The feature engineering module converts cleaned data into numerical features suitable for model input; one-hot encoding is used for unordered categorical variables; label encoding is used for ordered categorical variables; C. Feature screening: The feature screening module filters the numerical features converted by the feature engineering module to select the features that have the greatest impact on customer flow; D. Data Enhancement: Import population migration data into the data enhancement module. Based on the preset population migration ratio threshold, select cities with strong correlation with the target hub and construct a population in-migration / out-migration matrix. In this step, the immigration / emigration matrix includes an immigration matrix and an outgoing matrix; the immigration matrix represents the percentage of population migration from each city to the target transportation hub, and the outgoing matrix represents the percentage of population migration from the target transportation hub to each city.

[0049] In this step, the migration ratio threshold includes the in-migration ratio threshold and the out-migration ratio threshold. Meeting either of these two thresholds is deemed to meet the migration ratio threshold; the in-migration ratio threshold: the average daily population inflow from the city to the target hub accounts for ≥1% of the total inflow of the target hub; the out-migration ratio threshold: the average daily population outflow from the target hub to the city accounts for ≥1% of the total outflow of the target hub.

[0050] In this step, the feature screening operation is as follows: first, Spearman correlation analysis is used to calculate the correlation coefficient between each feature and passenger flow. Based on the rule that the absolute value of the Spearman correlation coefficient satisfies |ρ| ≥ 0.5, strongly correlated features are retained, and based on the rule that the absolute value of the Spearman correlation coefficient satisfies |ρ| < 0.3, weakly correlated features are eliminated. Then, a recursive feature elimination method is used. Based on the random forest model, the importance of features is evaluated by the decrease of the Gini index. The feature with the lowest importance ranking is eliminated in each iteration. The mean square error of the random forest model is monitored through 5-fold cross-validation. When the mean square error drops by more than 5%, the elimination is terminated and the top 10 features are retained.

[0051] In this step, the scheme for filling missing values ​​is as follows: for numerical features, if the missing ratio is less than 5%, the mean is used for filling; if the missing ratio is ≥5%, 0 is used for filling or interpolation of previous and next values; for categorical features, the mode is used for filling; for missing periods in time series features, linear interpolation or data from neighboring cities are used to replace them.

[0052] In this step, the outlier elimination scheme is as follows: outliers are eliminated by combining statistical methods and business rules; for numerical features, the 3σ principle is used to eliminate outliers, and then unreasonable data is eliminated by combining domain knowledge.

[0053] S03, lightweight CNN model construction: A. Frequency Domain Feature Extraction: Adaptive wavelet analysis module is used to implement the following operations: Wavelet decomposition: Adaptive wavelet function is used to decompose time domain passenger flow data into frequency domain components, extracting high-frequency and low-frequency features respectively; Frequency Domain Feature Meshing: Two independent one-dimensional convolutional neural networks are preset to fit the extracted high-frequency and low-frequency components respectively; Nonlinear Enhancement: The GELU operator is used to enhance the nonlinear expression capability of features; B. CNN model construction: The constructed CNN network module includes a one-dimensional convolutional layer and a parameter sharing strategy. The one-dimensional convolutional layer uses depthwise separable convolution to decompose the standard convolution into depthwise convolution and pointwise convolution. The parameter sharing strategy shares the convolution kernel weights across multiple time windows. C. Data fusion: The data fusion module is used to fuse the population in-migration / out-migration matrix with the frequency domain components in a weighted manner, thereby quantifying the contribution of population mobility to passenger flow fluctuations.

[0054] In this step, the network structure of the CNN network module includes an input layer, a pooling layer, and a fully connected layer in addition to a one-dimensional convolutional layer. The input layer is used to receive frequency domain components and the population in-migration / out-migration matrix. The pooling layer uses maximum pooling to compress feature dimensions and retain significant features. The number of output neurons in the fully connected layer is consistent with the prediction time span.

[0055] S04, model training and optimization: A. Model training: Pre-train the backbone network (i.e., the lightweight CNN model constructed in the previous step) based on historical passenger flow data to enable the model to initially learn the basic temporal patterns of passenger flow. The loss function uses mean squared error and the Adam optimizer is used. B. Model Optimization: The population in-migration / out-migration matrix and the multiple features with the greatest impact on passenger flow selected in step S02 are used as augmented data to retrain the pre-trained backbone network to optimize the model's ability to model inter-city interactions and external events.

[0056] In this step, during the second training process, the convolutional layer weights of the low-frequency components in the backbone network are fixed, and only the weights of the high-frequency components and the data enhancement channel are updated.

[0057] In this step, during the two training processes, the gradient of the loss function with respect to the model parameters is calculated by automatic differentiation, and the model weight parameters are iteratively updated based on the gradient direction.

[0058] S05, get the prediction result: A. Time Domain Signal Reconstruction: After weighted fusion of the population in-migration / out-migration matrix and the frequency domain components, a time domain prediction sequence is generated through inverse wavelet transform, and a sequence of passenger flow prediction values ​​for future time periods is output; B. Forecast result output: The result output module outputs the forecast results in JSON or CSV format, including the date, predicted passenger flow, and confidence interval.

[0059] In this step, when the population migration data and auxiliary data are updated, the result output module re-outputs the prediction results.

[0060] Verification of the technical effect of the present invention (taking the passenger flow forecast of Hangzhou East Station as an example): Dataset: Historical data from July 2016 to July 2022, divided into a training set (70%), a validation set (15%), and a test set (15%). The mean squared error (MSE) was calculated between the predicted results and the true values, and the prediction results of various models were compared. The comparison results are shown in Table 1, showing that the prediction algorithm in this paper achieves a smaller MSE loss value.

[0061] Table 1:

[0062] The "health incident" mentioned in this invention is the abbreviation of "public health incident".

Claims

1. The lightweight transportation hub passenger flow prediction system is characterized by: Includes multi-source data acquisition module, data processing and enhancement module and lightweight CNN model construction module; The multi-source data acquisition module is used to collect historical passenger flow data, population migration data and other auxiliary data; The data processing and enhancement module is used to process the raw data collected by the multivariate data acquisition module and output the training set, validation set and test set; the data processing and enhancement module includes a data cleaning module, a feature engineering module, a feature screening module, a data enhancement module and a data post-processing module; The data enhancement module combines population migration data with a preset population migration ratio threshold to screen cities that are strongly correlated with the target hub and construct a population in-migration / out-migration matrix based on this data. The lightweight CNN model construction module is used to realize passenger flow prediction; the lightweight CNN model construction module includes an adaptive wavelet analysis module, a CNN network module, a data fusion module, a time domain signal reconstruction module and a result output module; the data fusion module is used to weightedly fuse the population in / out matrix with the frequency domain component.

2. The lightweight transportation hub passenger flow prediction system according to claim 1 is characterized in that: In the data processing and enhancement module: the data cleaning module is used to process missing values ​​and outliers in the data; the feature engineering module is used to convert the data processed by the data cleaning module into numerical features suitable for model input; The feature screening module selects the features that have the greatest impact on passenger flow based on the numerical features converted by the feature engineering module. The data post-processing module normalizes the data processed by the data cleaning module, feature engineering module, feature screening module, and data enhancement module. It then divides the normalized data into training sets, validation sets, and test sets according to time windows, and finally converts the data into the model input format. In the lightweight CNN model construction module: the CNN network module is trained on the normalized data output by the data post-processing module; the network structure of the CNN network module includes a one-dimensional convolution layer; the one-dimensional convolution layer uses depthwise separable convolution to decompose the standard convolution into depthwise convolution and point-by-point convolution, reducing the number of parameters to 1 / N of the traditional convolution, where N is the number of channels; the time domain signal reconstruction module is used to map the weighted fused frequency domain signal back to the time domain through an inverse wavelet transform to form the final passenger flow prediction value; the result output module is used to output the passenger flow prediction result including the date, predicted passenger flow and confidence interval.

3. The lightweight transportation hub passenger flow prediction system according to claim 1 or 2, characterized in that: Other auxiliary data also include temperature data, air quality data and health-related data.

4. A lightweight transportation hub passenger flow prediction method, based on the lightweight transportation hub passenger flow prediction system according to claim 3, characterized in that the method as follows: S01, multi-source data acquisition: The following data is collected through the multi-source data collection module: historical passenger flow data: including the number of people arriving and departing from the station, train number and timestamp; population migration data: the percentage of population in / out between the transportation hub and other cities outside the transportation hub city; other auxiliary data: including weather data and holiday data; S02, data preprocessing and enhancement: A. Data cleaning: The data cleaning module processes missing values ​​and outliers in the collected multi-source data, fills in missing values, and removes outliers to ensure data integrity and reliability. B. Feature Engineering: The feature engineering module converts cleaned data into numerical features suitable for model input; one-hot encoding is used for unordered categorical variables; label encoding is used for ordered categorical variables; C. Feature screening: The feature screening module filters the numerical features converted by the feature engineering module to select the features that have the greatest impact on customer flow; D. Data Enhancement: Import population migration data into the data enhancement module. Based on the preset population migration ratio threshold, select cities with strong correlation with the target hub and construct a population in-migration / out-migration matrix. S03, lightweight CNN model construction: A. Frequency domain feature extraction: Adaptive wavelet analysis module is used to implement the following operations: Wavelet decomposition: Adaptive wavelet function is used to decompose the time domain passenger flow data into frequency domain components, and high-frequency and low-frequency features are extracted respectively; Frequency domain characteristics Meshing: Two independent one-dimensional convolutional neural networks are preset to fit the extracted high-frequency and low-frequency components respectively; Nonlinear enhancement: Enhance the nonlinear expression ability of features through the GELU operator; B. CNN model construction: The constructed CNN network module includes a one-dimensional convolutional layer and a parameter sharing strategy. The one-dimensional convolutional layer uses depthwise separable convolution to decompose the standard convolution into depthwise convolution and pointwise convolution. The parameter sharing strategy shares the convolution kernel weights across multiple time windows. C. Data fusion: The data fusion module is used to fuse the population in-migration / out-migration matrix with the frequency domain components in a weighted manner to quantify the contribution of population mobility to passenger flow fluctuations; S04, model training and optimization: A. Model training: The backbone network is pre-trained based on historical passenger flow data to enable the model to initially learn the basic temporal patterns of passenger flow. The loss function uses mean square error and the Adam optimizer is used. B. Model Optimization: The population in-migration / out-migration matrix and the multiple features with the greatest impact on passenger flow identified in step S02 are used as augmented data to retrain the pre-trained backbone network to optimize the model's ability to model inter-city interactions and external events. S05, get the prediction result: A. Time Domain Signal Reconstruction: After weighted fusion of the population in-migration / out-migration matrix and the frequency domain components, a time domain prediction sequence is generated through inverse wavelet transform, and a sequence of passenger flow prediction values ​​for future time periods is output; B. Forecast result output: The result output module outputs the forecast results in JSON or CSV format, including the date, predicted passenger flow, and confidence interval.

5. The lightweight transportation hub passenger flow prediction method according to claim 4, characterized in that: In step S01, weather data includes weather conditions and extreme weather events; temperature data includes daily average temperature and daily extreme temperature and corresponding time; holiday data includes the date and length of statutory holidays, as well as the date and length of special holidays.

6. The lightweight transportation hub passenger flow prediction method according to claim 5, characterized in that: In step S01, other auxiliary data also include temperature data, air quality data and health-related data; temperature data includes daily average temperature, daily temperature extremes and corresponding time; air quality data includes pollutant concentration values ​​and air quality levels; health-related data includes regional health risk levels and travel restrictions.

7. The lightweight transportation hub passenger flow prediction method according to claim 6, characterized in that: In step S02, the in-migration / out-migration matrix includes an in-migration matrix and an out-migration matrix; the in-migration matrix represents the percentage of population in-migration from each city to the target transportation hub, and the out-migration matrix represents the percentage of population out-migration from the target transportation hub to each city; in step S02, the migration ratio threshold includes an in-migration ratio threshold and an out-migration ratio threshold, and satisfying either of these two thresholds is deemed to satisfy the migration ratio threshold; in-migration ratio threshold: the average daily population in-migration from the city to the target hub accounts for ≥1% of the total in-migration of the target hub; out-migration ratio threshold: the average daily population out-migration from the target hub to the city accounts for ≥1% of the total out-migration of the target hub.

8. The lightweight transportation hub passenger flow prediction method according to claim 7, characterized in that: In step S02, the feature screening operation is as follows: first, the Spearman correlation analysis is used to calculate the correlation coefficient between each feature and passenger flow. Based on the rule that the absolute value of the Spearman correlation coefficient satisfies |ρ| ≥ 0.5, strongly correlated features are retained, and based on the rule that the absolute value of the Spearman correlation coefficient satisfies |ρ| < 0.3, weakly correlated features are eliminated. Then, a recursive feature elimination method was used. Based on the random forest model, the importance of features was evaluated by decreasing the Gini index. The feature with the lowest importance ranking was eliminated in each iteration. The mean square error of the random forest model was monitored through 5-fold cross-validation. When the mean square error dropped by more than 5%, the elimination was terminated and the top 10 features were retained.

9. The lightweight transportation hub passenger flow prediction method according to claim 8, characterized in that: In step S02, the scheme for filling missing values ​​is as follows: for numerical features, if the missing ratio is less than 5%, the mean is used for filling; if the missing ratio is ≥5%, 0 is used for filling or interpolation of previous and next values; for categorical features, the mode is used for filling; for missing periods in time series features, linear interpolation or data from neighboring cities are used to replace them; in step S02, the scheme for eliminating outliers is as follows: outliers are eliminated by combining statistical methods and business rules; for numerical features, the 3σ principle is used to eliminate outliers, and then unreasonable data is eliminated by combining domain knowledge.

10. The lightweight transportation hub passenger flow prediction method according to claim 9, characterized in that: In step S03, the network structure of the CNN network module includes an input layer, a pooling layer, and a fully connected layer in addition to a one-dimensional convolutional layer. The input layer is used to receive frequency domain components and the population in-migration / out-migration matrix. The pooling layer uses maximum pooling to compress feature dimensions and retain significant features. The number of output neurons in the fully connected layer is consistent with the prediction time span. In step S04, during the second training process, the convolutional layer weights of the low-frequency components in the backbone network are fixed, and only the weights of the high-frequency components and the data enhancement channel are updated; In step S04, during the two training processes, the gradient of the loss function with respect to the model parameters is calculated by automatic differentiation, and the model weight parameters are iteratively updated based on the gradient direction.

Images