A method and system for dynamic processing of intermodal ticketing based on load forecasting and implicit customers.

Abstract

This invention provides a method and system for dynamic processing of intermodal ticketing based on load forecasting and implicit customer data. It solves the technical obstacles that lead to inaccurate pricing decisions and user complaints due to reliance on a single, delayed, or erroneous external data source; the inability of fixed-discount intermodal tickets to adapt to real-time market supply and demand, resulting in revenue loss; the inability of discount strategies to differentiate customers during off-peak periods, leading to wasted discount resources; and the inability to quantify inventory data, resulting in a crude intermodal cooperation mechanism and difficulty in fairly measuring the contributions of all parties. This invention introduces a system load rate model for each time window, remapping any valid continuous pass scenario for any train into a series of discrete time windows for calculating expected load data. It achieves forward-looking perception of market demand and adaptive price generation, thereby realizing dynamic adjustment of fares, precise guidance of demand, and intelligent allocation of revenue; enabling the intermodal ticketing system to truly possess intelligent revenue management capabilities.

Description

Technical Field

[0001] This invention relates to the field of intermodal transport ticketing, and in particular to a method and system for dynamic processing of intermodal transport ticketing based on load forecasting and implicit customers. Background Technology

[0002] With the deepening integration of civil aviation and ground transportation, the "airfare + airport bus" intermodal travel model has become a core path to improve passenger experience and optimize operational revenue. The market's demand for intelligent ticketing systems that offer "one ticket for all, seamless connections" is increasingly urgent. However, mainstream technical solutions still remain at the stage of static discounts and independent ticketing, failing to achieve dynamic matching of supply and demand and intelligent revenue management.

[0003] In publicly available technologies, "fixed-discount multi-trip tickets" are commonly used. This model has inherent limitations: prices remain static and cannot automatically adjust to changes in market supply and demand. This leads to low prices and lost revenue during peak periods, and high prices and idle capacity during off-peak periods. The system fails to optimize resources or maximize revenue, limiting the commercial potential of multi-trip transportation. The root cause of not adopting dynamic multi-trip pricing is not insufficient algorithmic capability, but rather a fundamental conflict in the data structure between airport bus pass operation models and dynamic pricing algorithms. Existing dynamic pricing systems, such as airline tickets or hotel room rates, rely on discretizable inventory units, such as seats or rooms on specific dates, to perform quantitative calculations and price optimization. However, airport bus passes are "valid for any trip," and their "capacity inventory" is continuous and non-specific over time, presenting a fuzzy, dynamic, and unenumerable state. Such an inventory form cannot be recognized and processed as algorithmic input, rendering traditional dynamic pricing methods completely ineffective in multi-trip scenarios.

[0004] This structural incompatibility directly leads to three systemic defects in the intermodal transport system: First, the supply and demand signals are distorted, the system cannot quantify the potential passenger flow pressure at a certain time in the future, and prices cannot reflect the true market situation, resulting in both peak-season revenue loss and off-peak capacity waste; second, the pricing strategy is simplistic, due to the lack of calculable inventory indicators, only a fixed discount policy can be implemented, and it is impossible to conduct precise segmentation and differentiated discounts based on customer characteristics; third, the profit-sharing mechanism is rigid, and the fixed percentage sharing makes it difficult to reflect the dynamic contributions of airlines and bus companies at different stages, weakening the enthusiasm for cooperation and the fairness of revenue.

[0005] In summary, the key bottleneck of existing technologies lies in the lack of a fundamental method to transform the fuzzy passenger flow under a pass-through system into structured supply and demand indicators that algorithms can process. Without establishing this bridge, any intelligent pricing, revenue management, or resource scheduling algorithms will be unable to be implemented. Summary of the Invention

[0006] To address the aforementioned technical contradictions and business challenges, this invention proposes a novel technical solution. Its key feature is the introduction of a "time window system load rate" model, which remaps the continuous pass scenario of "valid for any number of trips" into a series of calculable expected load data within discrete time windows. In this way, a measurable digital twin is established for the fuzzy pass inventory, enabling it to be invoked and optimized in real time by the dynamic pricing engine. Furthermore, the system combines macro-level load forecasting with user intention modeling to achieve forward-looking perception of market demand and adaptive price generation, thereby realizing dynamic adjustment of fares, precise guidance of demand, and intelligent allocation of revenue. The technical solution of this invention fundamentally solves the long-standing technical bottleneck of "uncalculable pass inventory" in the industry, enabling intermodal ticketing systems to truly possess intelligent revenue management capabilities.

[0007] This solution addresses the technical hurdles that lead to inaccurate pricing decisions and even user complaints due to reliance on a single, delayed, or erroneous external data source; overcomes the shortcomings of traditional fixed-discount tickets, which cannot adapt to real-time market supply and demand changes, resulting in lost revenue during peak periods and idle seats during off-peak periods; resolves the fundamental business pain point that traditional discount strategies fail to differentiate customers, wasting significant marketing resources on "existing customers" who will consume regardless of circumstances, thus eroding net profit; addresses the fundamental contradiction between the "quantified inventory" data required for dynamic pricing algorithms and the actual operating model of bus tickets, which are "not sold by seat or even by trip (valid for any trip)"; and resolves the business challenge of long-term stable cooperation in intermodal transport collaborations, where a crude revenue distribution mechanism makes it difficult to fairly measure the contributions of all parties.

[0008] This invention provides a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers. The steps of this method are as follows: Step 1: Data Acquisition and Fusion In each time window Within this process, raw data is acquired, including data from K flights and real-time data from M external business sources; and data initialization is completed within a preset time to obtain a confidence feature vector. ; Step 2: Calculation of system load rate within the time window Confidence feature vector The data is fed into the load prediction model (XGBoost) to obtain the time window. Total demand Total available transport capacity And calculate the load rate. ; Step 3: Customer Intent Prediction and Implicit Segmentation Get Time Window Any candidate passenger eigenvectors ; will passengers Features Input into the wide-depth hybrid prediction model, the model outputs passenger data. Probability of immediate purchase given a price. ; Obtain passenger Price sensitivity and rigid demand, and will passengers Automatically categorize customers and assign them a discrete label. The subsequent pricing is guided by the segmentation logic; Step 4: Dynamic Pricing and Demand Processing Based on real-time system load rate With passengers Purchase probability distribution Automatic solution time window Best fare or discount level The discount must satisfy load constraints, regulatory constraints, and price smoothing constraints; Step 5: Revenue Sharing and Operational Collaboration Passengers After a successful transaction, a two-step pruning approximate Shapley value algorithm is used to calculate and attribute the revenue to each party in real time; when the future time window... The system load prediction meets the continuous load condition. At that time, the system generates a capacity suggestion vector and sends an early warning to the operations dispatch center.

[0009] Preferably, the data initialization in step 1 involves identifying and repairing the quality of the original data; specifically as follows: Step 1.1: Data initialization begins by determining if the source data falls within the time window. Whether it is trustworthy, the system determines the reliability of each data source. Calculate credibility score Defined as: (1) in, The timeliness of the data is defined as follows: , For source The latest data update time is far from the time window Time difference (minutes) The time decay constant; Represented as a time window The first Data sources; This indicates the consistency between each data source and the system reference baseline; it is defined as: (2) in, It is the historical average, the weighted average of the same indicator from other sources, or the smoothed value of the previous time window; Indicate source In the time window The original data values, To prevent tiny constants with a denominator of 0; , For the weighting coefficients, satisfying + =1; Step 1.2: Identification and repair of raw data quality is performed by the system using a global threshold. Examine the data quality from each source: like If so, the data from that source is considered reliable; like If the data is unreliable, it should be corrected. For low-reliability data, the system... Perform a fast correction; the correction method is a weighted average with historical smoothed values. (3) in, It is a smoothed reference value from the previous time window or the same time period of the previous day; The values ​​are set based on business importance and scenario reliability; after correction Only then will it be adopted; Step 1.3: The system concatenates reliable or corrected data from all sources to obtain a time window. Confidence feature vector All sources refer to the system within the time window. The data sub-streams obtained from external modules include flight data, bus data, ticketing data, and weather data; Data concatenation involves linking the data vectors from different sources end to end to form a larger input vector. (4).

[0010] Preferably, the time window in step 2 System load rate calculation is to use vectors Directly input into the load forecasting model to obtain the forecast time window. Total demand Total available transport capacity and load rate , Define time window The time interval covered is: , Indicates the duration of the window, from which the window begins to open. Beginning, to Finish; Define time window Available transport capacity for: (5) in, For time window Number of departures This refers to the number of seats per vehicle. The duration of the time window (in hours); Time window based on machine learning model Demand forecast Defined as: (6) Where A is the confidence feature vector output in step 1. Describe the time window Environment state; function It is a trained prediction model that has the ability to process multidimensional nonlinear relationships, heterogeneous features and time series data. The model can be selected from gradient boosting tree model (XGBoost), random forest, multilayer perceptron, deep neural network, long short-term memory network or gated recurrent unit. Time window The load rate is: (7).

[0011] Preferably, the ticket-buying passenger in step 3 eigenvectors for: (8) in, This refers to user profiles, with typical fields including age, place of residence, travel frequency, and average advance ticket purchase time. Refers to the time window Environmental changes, including holidays, weather, and flight density; Time window Internal load rate; This refers to the compatibility between a user's historical travel routes and their target route. Refers to the most recent n ticket purchase records and payment channels; eigenvectors The input is fed into a lightweight, wide-depth mixed prediction model; the model outputs the instantaneous purchase probability of the passenger given the price, denoted as: (9) in, Indicates passenger In the time window The event of completing the ticket purchase within the specified time; It refers to a specific candidate discount or offer within the system; Then inform the model of this tourist Who is it, what is the current situation, and what price are we offering? Based on real-time purchase probability Passengers The system automatically categorizes customers into three types: existing customers (those with essential needs), wavering customers (price-sensitive), and low-interest customers (those with weak relevance); and provides services for travelers. In the time window Assign a discrete label ; Customers are categorized as follows: Existing customers (those with essential needs): If they still have a high probability of making a purchase even without discounts: Then it is believed that the passenger He will basically buy unconditionally, so discounts have little impact on him. The value is taken from experience. Swing customers (price-sensitive): They are less likely to buy when there is no discount, but their purchase probability increases significantly when there is a moderate discount. ; in, It is a threshold for discount-driven volume increase; Low-intention customers (weakly correlated): These are customers who, even with a high discount, are still less likely to purchase than a certain threshold. ; in, , , Thresholds for classifying customer preferences; The threshold for increasing purchase volume driven by discounts.

[0012] Preferably, the system solves for the time window in step 4. Discounts within The discount ceiling is defined using the Sigmoid smooth discount adjustment function: (10) in, Indicates the maximum allowed discount; For the Sigmoid function; Indicates adjustment of steepness; Indicates the center point of the Sigmoid function; Let the predicted demand within the time window be... Available transport capacity is The upper limit of potential ticket purchasers For any candidate discount The corresponding bus fare is: ,in, The base fare; In the time window The average purchase probability of the target passenger group is: (11) The expected number of actual transactions is: (12) Discount The expected return is: (13) in, It is a cost function related to discounts, including discount concessions and channel promotion expenses; the cost function related to discounts is defined as: , This is the discount cost coefficient; The system finds the optimal discount level by searching (or analytically approximating) within the allowed interval.

[0013] (14) Simultaneously satisfying the constraints

[0014] in, The minimum allowed ticket price, The maximum magnitude of the two price adjustments, Discount for the previous window.

[0015] The allowable interval refers to a feasible discount region jointly defined by constraints.

[0016] Preferably, the real-time calculation and attribution of benefits to each party in step 5 is as follows: Step 5.1: Settlement period is a time window The intermodal transport transactions actually completed within the region constitute a distributable revenue pool, and the distributable net revenue of the revenue pool is... Let the set of partners who participated in this intermodal transport service and contributed to the transaction be: The value function of the participants is defined as follows: ,in, It is a subset of any set of participants. Indicates only the participating subset Total revenue that can be obtained when participating in intermodal transport services; Step 5.2: Classic Shapley values ​​for each participant Define the weighted average of its marginal contributions: (15) When the Fang joins any existing group of partners What is the expected increase in revenue when adding the item, and what is the expected value for all possible order of addition? The order in which participants join the system is random, and the contribution of the same participant will change depending on the order of joining. It is necessary to calculate the expectation over all joining orders, that is, to calculate the average value. Step 5.3: The two-step pruning approximation Shapley value algorithm consists of a contribution pruning algorithm and a federation value pruning algorithm, the details of which are as follows: Step 5.3.1: The system employs contribution-based pruning, first estimating the upper bound of the marginal contribution of each participant. : (16) in, This represents the potential revenue generated when all stakeholders in the system participate. Indicates the removal of the main body The benefits that the system can generate; Then the participating parties are considered to be within the time window. The marginal effect of those involved is minimal and they are not included in the real-time amortization calculation set, resulting in a reduced set of core participants: (17) in, The contribution threshold; Step 5.3.2: Gathering core stakeholders For the remaining few core participants, the system uses Monte Carlo sampling to estimate the Shapley value, obtaining an approximate Shapley contribution for each retainer, denoted as . Ultimately, each retainer was obtained. Weighting allocation: (18) According to the time window Distributable net income of the income pool Divide proportionally: (19) in, The participants are within the time window The corresponding settlement amount; Step 5.4: After revenue sharing, the system plans for the future. The load rate prediction is calculated for each consecutive time window:

[0017] Set a capacity strain threshold When satisfied It is a natural number, meaning the system predicts that the system will continue to operate at high load; the system generates capacity intervention suggestions. Define the suggestion vector as follows: (20) in, For the suggested increase in the number of classes, For the suggested increase in seating capacity, High-risk routes are numbered. The specific time window for the proposed addition.

[0018] This invention provides a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customer data. It addresses the technical obstacles caused by reliance on a single, delayed, or erroneous external data source, leading to inaccurate pricing decisions and even user complaints. It overcomes the shortcomings of traditional fixed-discount intermodal tickets, which cannot adapt to real-time market supply and demand changes, resulting in lost revenue during peak periods and idle seats during off-peak periods. It also solves the problem of traditional discount strategies failing to differentiate customers, wasting significant marketing resources on "existing customers" who will consume regardless of circumstances. Furthermore, it addresses the need for quantified inventory data in dynamic pricing algorithms. In intermodal cooperation, the crude revenue distribution mechanism makes it difficult to fairly measure the contributions of all parties. This invention introduces a time-window system load rate model, remapping any valid continuous pass scenario for any train into a series of calculable expected load data within discrete time windows. It establishes a measurable digital twin, enabling real-time calling and optimization by the dynamic pricing engine. The system combines macro-level load forecasting with user intention modeling to achieve forward-looking perception of market demand and adaptive price generation, thereby realizing dynamic adjustment of fares, precise guidance of demand, and intelligent allocation of revenue. This truly endows the intermodal ticketing system with intelligent revenue management capabilities.

[0019] In one embodiment, and in another aspect, the present invention provides a system for dynamic processing of intermodal ticketing based on load forecasting and implicit customers. This system includes a data acquisition and fusion module, a load forecasting module, a customer segmentation and willingness forecasting module, a dynamic pricing and demand processing module, and a revenue sharing and operational coordination module. Data acquisition and fusion module: Used for multi-source heterogeneous data access, cleaning, and fusion; within each time window Within this process, raw data is acquired, including real-time data from external business sources for K flights and M flights; and data initialization is completed within a preset time to obtain a confidence feature vector. ; Load forecasting module: Used for macro-level passenger flow forecasting based on machine learning; defines time windows. , confidence feature vector Input the data into the load forecasting model (XGBoost) to obtain the forecast time window. Total demand Total available transport capacity and load rate ; Customer segmentation and intention prediction module: Used for micro-level modeling of individual customer purchase intentions; obtaining time windows. Any candidate passenger eigenvectors ; will passengers Features Input into the wide-depth hybrid prediction model, the model outputs passenger data. Probability of immediate purchase given a price. ; Obtain passenger Price sensitivity and rigid demand, and will passengers Classify customers into different types and assign them a discrete label. The subsequent pricing is guided by the segmentation logic; Dynamic pricing and demand processing module: A decision engine that combines load rate and willingness probability; based on real-time system load rate. With passengers Purchase probability distribution Automatic solution time window Best fare or discount level ; Revenue Sharing and Operational Coordination Module: Used for revenue sharing and dynamic capacity scheduling; Passengers After a transaction is successful, a two-step pruning approximate Shapley value algorithm is used to calculate and attribute the benefits to each party in real time. When the future time window The system load prediction meets the continuous load condition. At that time, the system generates a capacity suggestion vector and sends an early warning to the operations dispatch center.

[0020] Preferably, the data acquisition and fusion module includes a raw data credibility scoring module, a raw data quality identification module, and a raw data repair module. The system assigns a raw data credibility score to determine the reliability of source data within a given time window. This score is applied to each data source. Calculate credibility score Defined as: (1) in, The timeliness of the data is defined as follows: , Indicated as source The latest data update time is far from the time window Time difference (minutes) The time decay constant; Represented as the first within the time window Data sources; This indicates the consistency between each data source and the system reference baseline; it is defined as: (2), in, It is the historical average, the weighted average of the same indicator from other sources, or the previous time window. Smoothing values ​​within; To prevent tiny constants with a denominator of 0; , For the weighting coefficients, satisfying + =1; The raw data quality assessment module uses a global threshold. Examine the data quality from each source, if If so, the data from that source is considered reliable; like If the data is unreliable, it should be corrected. For low-reliability data, the system will... Perform a fast correction; the correction method is a weighted average with historical smoothed values. (3) in, It is a smoothed reference value from the previous time window or the same time period of the previous day; The value is set based on business importance and scenario reliability; ; The raw data repair module is used to stitch together reliable or corrected data from all sources to obtain a time window. Confidence feature vector All sources refer to the system within the time window. The data sub-streams obtained from external modules include flight data, bus data, ticketing data, and weather data; Data concatenation, simply put, is the process of linking the data vectors from different sources together to form a larger input vector. (4).

[0021] Preferably, the prediction time window is obtained through a load prediction model. Total demand Total available transport capacity and load rate The specific details are as follows: Define time window The time interval covered is: , Indicates the duration of the window, from which the window begins to open. Beginning, to Finish; Define time window Available transport capacity for: (5) in, For time window Number of departures This refers to the number of seats per vehicle. The duration of the time window (in hours); Time window based on machine learning model Demand forecast Defined as: (6) Where A is the confidence feature vector output in step 1. Describe the time window Environment state; function It is a trained prediction model that has the ability to process multidimensional nonlinear relationships, heterogeneous features and time series data. The model can be selected from gradient boosting tree model (XGBoost), random forest, multilayer perceptron, deep neural network, long short-term memory network or gated recurrent unit. Time window The load rate is: (7) The customer segmentation and willingness prediction module includes a customer segmentation module and a willingness prediction module; Customer segmentation module, based on ticket-purchasing passengers eigenvectors Input into a lightweight wide-depth mixture prediction model, ticket-buying passengers eigenvectors for: (8) in, This refers to user profiles, with typical fields including age, place of residence, travel frequency, and average advance ticket purchase time. Refers to the time window Environmental changes, such as holidays, weather, and flight frequency; Time window Internal load rate; This refers to the compatibility between a user's historical travel routes and their target route. Refers to the most recent n ticket purchase records and payment channels; The probability of the passenger making an immediate purchase under a given price condition, output by the wide-depth hybrid prediction model, is denoted as: (9) in, Indicates passenger In the time window The event of completing the ticket purchase within the specified time; It is a candidate discount or offer level in the system; F𝑢( Then inform the model of this tourist. Who is it, what is the current situation, and what price are we offering? Based on real-time purchase probability Passengers The system automatically categorizes customers into three types: existing customers, wavering customers, and low-interest customers; and provides services for travelers. In the time window Assign a discrete label ; The willingness prediction module makes predictions based on customer categories: Existing customers still have a high probability of making a purchase even without discounts: , Taking experience as an example, it is believed that passengers They will generally buy unconditionally, so discounts have little impact on them; their existing customers are those with essential needs. Swing customers are less likely to buy when there is no discount, but their purchase probability increases significantly when there is a moderate discount: , in Discounts are the threshold for driving volume increases. Swing customers are price-sensitive customers; Low-intention customers: These are customers who, even with significant discounts, are still less likely to purchase than a certain threshold. Low-willingness customers are weakly related customers.

[0022] in , , Thresholds for classifying customer preferences; The threshold for increasing purchase volume driven by discounts.

[0023] Preferably, the dynamic pricing and demand processing module further includes an environment status input unit, an action selection unit, an execution and feedback unit, a reward and benefit unit, and a strategy update unit; The environment status input unit is used to input status information including system load rate, customer profile vector, and time information; system load rate Based on the predicted demand With total available capacity Calculations show that The target passenger group profile aggregation vector is generated by statistically analyzing the characteristics of all passengers using an aggregation function. Time characteristics include hourly features, holidays, and flight peaks; ( In ) The system also utilizes the load rate to construct a discount cap constraint function (i.e., a Sigmoid smooth discount adjustment function): (10) in, Indicates the maximum allowed discount; For the Sigmoid function; Indicates adjustment of steepness; This represents the center point of the Sigmoid function; the upper limit of the discount is used as "part of the environment state" to constrain the agent's action set.

[0024] The action selection unit is the agent's choice of a discount action from a predefined discount set {0%, 10%, 20%, 30%}. ; The selected action must meet the following requirements: For candidate discounts, the discount must be within the allowed range (i.e., the feasible discount range). The corresponding bus fare is calculated as follows: ,in, This is the base fare.

[0025] The execution and feedback unit is where the system quotes a price based on the selected discount, obtains the user's purchase result and actual revenue; the probability of a single passenger purchasing is... During the time window The average purchase probability for domestic travelers is: (11) The expected number of actual transactions is: (12) Discount The expected return is: (13) in, This is a cost function related to discounts. Discount-related costs include discount concessions and channel promotion expenses, defined as: , This is the discount cost coefficient; the actual and expected returns will be entered into the reward unit. The reward unit calculates the reward value based on the difference between the actual reward and the maximum reward: ,in, This represents the baseline expected return with or without a discount; reward. Used to measure actions The advantages and disadvantages; The policy update unit is where the agent updates its policy parameters using Q-learning or DQN algorithms to maximize long-term returns. In the reinforcement learning DQN framework, the system finds the optimal discount level by searching or parsing an approximation within an allowed interval.

[0026] (14) Simultaneously satisfying the constraints

[0027] in, The minimum allowed ticket price, The maximum magnitude of the two price adjustments, The allowable interval is a feasible discount range defined by the constraints, representing the discount for the previous time window. The revenue sharing and operational coordination module includes a revenue sharing unit and an operational coordination unit; Revenue sharing unit, used for real-time calculation and attribution to all parties; settlement period is a time window. The intermodal transport transactions actually completed within the region constitute a distributable revenue pool, and the distributable net revenue of the revenue pool is... Let the set of partners who participated in this intermodal transport service and contributed to the transaction be: The value function of the participants is defined as follows: ,in, It is a subset of any set of participants. This indicates if only this group of participants is collaborating; the classic Shapley value represents each participant. Define the weighted average of its marginal contributions: (15) When the Fang joins any existing group of partners The expected value of the gains brought by the addition of all participants in the system is calculated over all possible joining orders. The order in which all participants join the system is random, and the contribution of the same participant will change with different joining orders. Therefore, it is necessary to calculate the expected value over all joining orders, i.e., to calculate the average value. A two-step pruning approximation Shapley value algorithm is used as both the contribution pruning algorithm and the alliance value pruning algorithm. The algorithm content is as follows: The system employs contribution-based pruning, first estimating the upper bound of the marginal contribution of each participant. : (16) in, This represents the potential revenue generated when all stakeholders in the system participate. This represents the revenue that the system can generate after removing subject i.

[0028] Then the participating parties are considered to be within the time window. The marginal effect of those involved is minimal and they are not included in the real-time amortization calculation set, resulting in a reduced set of core participants: (17) in, The contribution threshold; In the core stakeholder group For the remaining few core participants, the system uses Monte Carlo sampling to estimate the Shapley value, obtaining an approximate Shapley contribution for each retainer, denoted as . Ultimately, each retainer was obtained. Weighting allocation: (18) According to the time window Distributable net income of the income pool Divide proportionally: (19) in, The participants are within the time window The corresponding settlement amount; The operational coordination unit is used by the system to plan for future revenue sharing. The load rate prediction is calculated for each consecutive time window:

[0029] Set a capacity strain threshold When satisfied , It is a natural number, meaning the system's predicted value remains under high load; the system generates capacity intervention suggestions, defined as a suggestion vector: (20) in, For the suggested increase in the number of classes, For the suggested increase in seating capacity, High-risk routes are numbered. Specific time windows for the proposed addition When the system predicts that the load rate exceeds the threshold At that time, the system generates a capacity suggestion vector and sends an early warning to the operations dispatch center.

[0030] The advantages of this invention are: This invention provides a system for dynamic processing of intermodal ticketing based on load forecasting and implicit customer data. It addresses the technical obstacles caused by reliance on a single, delayed, or erroneous external data source, leading to inaccurate pricing decisions and even user complaints. It overcomes the shortcomings of traditional fixed-discount intermodal tickets, which cannot adapt to real-time market supply and demand changes, resulting in lost revenue during peak periods and idle seats during off-peak periods. It also solves the problem of traditional discount strategies failing to differentiate customers, wasting significant marketing resources on "existing customers" who will consume regardless of circumstances. Furthermore, it addresses the need for quantified inventory data in dynamic pricing algorithms, as the crude revenue distribution mechanism in intermodal cooperation makes it difficult to fairly measure the contributions of all parties. This invention introduces a time-window system load rate model, remapping any valid continuous pass scenario for any train into a series of calculable expected load data within discrete time windows. It establishes a measurable digital twin, enabling real-time calling and optimization by the dynamic pricing engine. The system combines macro-level load forecasting with user intention modeling to achieve forward-looking perception of market demand and adaptive price generation, thereby realizing dynamic adjustment of fares, precise guidance of demand, and intelligent allocation of revenue. This truly endows the intermodal ticketing system with intelligent revenue management capabilities. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers. Figure 2 This is a flowchart of an embodiment of a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers; Figure 3 This is a pricing flowchart of an embodiment of a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers; Figure 4 This is a system module diagram based on load forecasting and dynamic processing of intermodal ticketing for implicit customers. Detailed Implementation

[0032] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems and methods consistent with some aspects of the invention as detailed in the appended claims.

[0033] This invention provides a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers. The steps of this method are as follows: like Figure 1 As shown, a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers is described, and the steps of the method are as follows: Step 1: Data Acquisition and Fusion In each time window Within this process, raw data is acquired, including data from K flights and real-time data from M external business sources; and data initialization is completed within a preset time to obtain a confidence feature vector. ; Step 2: Calculation of system load rate within the time window Confidence feature vector The data is fed into the load prediction model (XGBoost) to obtain the time window. Total demand Total available transport capacity And calculate the load rate. ; Step 3: Customer Intent Prediction and Implicit Segmentation Get Time Window Any candidate passenger eigenvectors ; will passengers Features Input into the wide-depth hybrid prediction model, the model outputs passenger data. Probability of immediate purchase given a price. ; Obtain passenger Price sensitivity and rigid demand, and will passengers Automatically categorize customers and assign them a discrete label. The subsequent pricing is guided by the segmentation logic; Step 4: Dynamic Pricing and Demand Processing Based on real-time system load rate With passengers Purchase probability distribution Automatic solution time window Best fare or discount level The discount must satisfy load constraints, regulatory constraints, and price smoothing constraints; Step 5: Revenue Sharing and Operational Collaboration Passengers After a successful transaction, a two-step pruning approximate Shapley value algorithm is used to calculate and attribute the revenue to each party in real time; when the future time window... The system load prediction meets the continuous load condition. At that time, the system generates a capacity suggestion vector and sends an early warning to the operations dispatch center.

[0034] In one embodiment, the data initialization in step 1 involves identifying and repairing the quality of the original data; specifically as follows: Step 1.1: Data initialization begins by determining if the source data falls within the time window. Whether it is trustworthy, the system determines the reliability of each data source. Calculate credibility score Defined as: (1) in, The timeliness of the data is defined as follows: , Indicated as source The latest data update time is far from the time window Time difference (minutes) The time decay constant; Represented as a time window The first Data sources; This indicates the consistency between each data source and the system reference baseline; it is defined as: (2) in, It is the historical average, the weighted average of the same indicator from other sources, or the smoothed value of the previous time window; Indicate source In the window The original data values, To prevent tiny constants with a denominator of 0; , For the weighting coefficients, satisfying + =1; Step 1.2: Identification and repair of raw data quality is performed by the system using a global threshold. Examine the data quality from each source: like If so, the data from that source is considered reliable; like If the data is unreliable, it should be corrected. For low-reliability data, the system... Perform a fast correction; the correction method is a weighted average with historical smoothed values. (3) in, It is a smoothed reference value of the target data source in the previous time window or the same time period of the previous day; The values ​​are set based on business importance and scenario reliability; after correction Only then will it be adopted; Step 1.3: The system stitches together reliable or corrected data from all sources to obtain a time window. Confidence feature vector All sources refer to the system within the time window. The data sub-streams obtained from external modules include flight data, bus data, ticketing data, and weather data; Data concatenation involves linking the data vectors from different sources end to end to form a larger input vector. (4).

[0035] In one embodiment, the time window in step 2 System load rate calculation involves using the confidence feature vector. Directly input into the load forecasting model to obtain the forecast time window. Total demand Total available transport capacity and load rate , Define time window The time interval covered is: , Indicates the duration of the window, from which the window begins to open. Beginning, to Finish; Define time window Available transport capacity for: (5) in, The number of departures within a time window. This refers to the number of seats per vehicle. The duration of the time window (in hours); Time window based on machine learning model Demand forecast Defined as: (6) Where A is the confidence feature vector output in step 1. Describe the time window Environment state; function It is a trained prediction model that has the ability to process multidimensional nonlinear relationships, heterogeneous features and time series data. The model can be selected from gradient boosting tree model (XGBoost), random forest, multilayer perceptron, deep neural network, long short-term memory network or gated recurrent unit. Time window The load rate is: (7).

[0036] In one embodiment, the ticket-buying passenger in step 3 eigenvectors for: (8) in, This refers to user profiles, with typical fields including age, place of residence, travel frequency, and average advance ticket purchase time. Refers to the time window Environmental changes, including holidays, weather, and flight density; Time window Internal load rate; This refers to the compatibility between a user's historical travel routes and their target route. Refers to the most recent n ticket purchase records and payment channels; eigenvectors The input is fed into a lightweight, wide-depth mixed prediction model; the model outputs the instantaneous purchase probability of the passenger given the price, denoted as: (9) in, Indicates passenger In the time window The event of completing the ticket purchase within the specified time; It refers to a specific candidate discount or offer within the system; Then inform the model of this tourist Who is it, what is the current situation, and what price are we offering? Based on real-time purchase probability Passengers The system automatically categorizes customers into three types: existing customers (those with essential needs), wavering customers (price-sensitive), and low-interest customers (those with weak relevance); and provides services for travelers. In the time window Assign a discrete label ; Customers are categorized as follows: Existing customers (those with essential needs): If they still have a high probability of making a purchase even without discounts: Then it is believed that the passenger He will basically buy unconditionally, so discounts have little impact on him. The value is taken from experience. Swing customers (price-sensitive): They are less likely to buy when there is no discount, but their purchase probability increases significantly when there is a moderate discount. ; in, It is a threshold for discount-driven volume increase; Low-intention customers (weakly correlated): These are customers who, even with a high discount, are still less likely to purchase than a certain threshold. ; in , , Thresholds for classifying customer preferences; The threshold for increasing purchase volume driven by discounts.

[0037] In one embodiment, the system solves for the time window in step 4. Discounts within The discount ceiling is defined using the Sigmoid smooth discount adjustment function: (10) in, Indicates the maximum allowed discount; For the Sigmoid function; Indicates adjustment of steepness; Indicates the center point of the Sigmoid function; Set time window Internally predicted demand is Available transport capacity is The upper limit of potential ticket purchasers For any candidate discount The corresponding bus fare is: ,in, The base fare; In the time window The average purchase probability of the target passenger group is: (11) The expected number of actual transactions is: (12) Discount The expected return is: (13) in, It is a cost function related to discounts, including discount concessions and channel promotion expenses; the cost function related to discounts is defined as: , This is the discount cost coefficient; The system finds the optimal discount level by searching (or analytically approximating) within the allowed interval.

[0038] (14) Simultaneously satisfying the constraints

[0039] in, The minimum allowed ticket price, The maximum magnitude of the two price adjustments, For the previous time window Discount.

[0040] The allowable interval refers to a feasible discount region jointly defined by constraints.

[0041] In one embodiment, the real-time calculation and attribution of benefits to each party in step 5 is as follows: Step 5.1: Settlement period is a time window The intermodal transport transactions actually completed within the region constitute a distributable revenue pool, and the distributable net revenue of the revenue pool is... Let the set of partners who participated in this intermodal transport service and contributed to the transaction be: The value function of the participants is defined as follows: ,in, It is a subset of any set of participants. Indicates only the participating subset Total revenue that can be obtained when participating in intermodal transport services; Step 5.2 Classic Shapley value for each participant Define the weighted average of its marginal contributions: (15) When the Fang joins any existing group of partners What is the expected increase in revenue when adding the item, and what is the expected value for all possible order of addition? The order in which all participants join the system is random. The contribution of the same participant will change depending on the order in which they join. It is necessary to calculate the expectation of all joining orders, that is, to calculate the average value. Step 5.3: The two-step pruning approximation Shapley value algorithm consists of a contribution pruning algorithm and a federation value pruning algorithm, the details of which are as follows: Step 5.3.1: The system employs contribution-based pruning, first estimating the upper bound of the marginal contribution of each participant. : (16) in, This represents the potential revenue generated when all stakeholders in the system participate. Indicates the removal of the main body The benefits that the system can generate; Then the participating parties are considered to be within the time window. The marginal effect of those involved is minimal and they are not included in the real-time amortization calculation set, resulting in a reduced set of core participants: (17) in, The contribution threshold; Step 5.3.2: Gathering core stakeholders For the remaining few core participants, the system uses Monte Carlo sampling to estimate the Shapley value, obtaining an approximate Shapley contribution for each retainer, denoted as . Ultimately, each retainer was obtained. Weighting allocation: (18) According to the time window Distributable net income of the income pool Divide proportionally: (19) in, The participants are within the time window The corresponding settlement amount; Step 5.4: After revenue sharing, the system plans for the future. The load rate prediction is calculated for each consecutive time window:

[0042] Set a capacity strain threshold When satisfied , It is a natural number, meaning the system predicts that the system will continue to operate at high load; the system generates capacity intervention suggestions. Define the suggestion vector as follows: (20) in, For the suggested increase in the number of classes, For the suggested increase in seating capacity, High-risk routes are numbered. The specific time window for the proposed addition.

[0043] This embodiment provides a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customer data. It addresses the technical obstacles that lead to inaccurate pricing decisions and even user complaints due to reliance on a single, delayed, or erroneous external data source; overcomes the shortcomings of traditional fixed-discount intermodal tickets, which cannot adapt to real-time market supply and demand changes, resulting in lost revenue during peak periods and idle seats during off-peak periods; solves the problem that traditional discount strategies cannot differentiate customers, wasting significant marketing resources on "existing customers" who will consume regardless; and addresses the need for quantified inventory data in dynamic pricing algorithms. In intermodal cooperation, the crude revenue distribution mechanism makes it difficult to fairly measure the contributions of all parties. This invention introduces a time-window system load rate model, remapping any valid continuous pass scenario for any train into a series of calculable expected load data within discrete time windows; establishes a measurable digital twin, enabling real-time calling and optimization by the dynamic pricing engine; and combines macro-load forecasting with user intention modeling to achieve forward-looking perception of market demand and adaptive price generation, thereby realizing dynamic adjustment of fares, precise guidance of demand, and intelligent allocation of revenue. This truly endows the intermodal ticketing system with intelligent revenue management capabilities.

[0044] Another aspect of the specific embodiments of the present invention provides a system for dynamic processing of intermodal ticketing based on load forecasting and implicit customers, such as... Figure 4 As shown; The system includes a data acquisition and fusion module, a load forecasting module, a customer segmentation and willingness forecasting module, a dynamic pricing and demand processing module, and a revenue sharing and operational collaboration module.

[0045] Data acquisition and fusion module: used for multi-source heterogeneous data access, cleaning, and fusion; within each time window Within this process, raw data is acquired, including real-time data from external business sources for K flights and M flights; and data initialization is completed within a preset time to obtain a confidence feature vector. ; Load forecasting module: used for macro-level passenger flow forecasting based on machine learning; defines time windows. , confidence feature vector Input the data into the load forecasting model (XGBoost) to obtain the forecast time window. Total demand Total available transport capacity and load rate ; Customer segmentation and intention prediction module: used for micro-level modeling of individual customer purchase intentions; obtaining time windows. Any candidate passenger eigenvectors ; will passengers Features Input into the wide-depth hybrid prediction model, the model outputs passenger data. Probability of immediate purchase given a price. ; Obtain passengers' price sensitivity and demand rigidity, and then... Classify customers into different types and assign them a discrete label. The subsequent pricing is guided by the segmentation logic; Dynamic pricing and demand processing module: A decision engine that combines load rate and willingness probability; based on real-time system load rate. With passengers Purchase probability distribution Automatic solution time window Best fare or discount level ; Revenue Sharing and Operational Coordination Module: Used for revenue sharing and dynamic capacity scheduling; Passengers After a transaction is successful, a two-step pruning approximate Shapley value algorithm is used to calculate and attribute the revenue to each party in real time. When the future time window The system load prediction meets the continuous load condition. At that time, the system generates a capacity suggestion vector and sends an early warning to the operations dispatch center.

[0046] In one embodiment, the data acquisition and fusion module includes a raw data credibility scoring module, a raw data quality identification module, and a raw data repair module. The system assigns a raw data credibility score to determine the reliability of source data within a given time window. This score is applied to each data source. Calculate credibility score Defined as: (1) in, The timeliness of the data is defined as follows: , Indicated as source The latest data update time is far from the time window Time difference (minutes) The time decay constant; Represented as a time window The first Data sources; This indicates the consistency between each data source and the system reference baseline; it is defined as: (2) in, It is the historical average, the weighted average of the same indicator from other sources, or the previous time window. The smoothed value; To prevent tiny constants with a denominator of 0; , For the weighting coefficients, satisfying + =1; The raw data quality assessment module uses a global threshold. Examine the data quality from each source, if If so, the data from that source is considered reliable; like If the data is unreliable, it should be corrected. For low-reliability data, the system will... Perform a fast correction; the correction method is a weighted average with historical smoothed values. (3) in, It is a smoothed reference value of the target data source in the previous time window or the same time period of the previous day; The value is set based on business importance and scenario reliability;

[0047] The raw data repair module is used to stitch together reliable or corrected data from all sources to obtain a time window. Confidence feature vector All sources refer to the data sub-streams that the system obtains from external modules within the time window, including flight data, bus data, ticketing data, and weather data; Data concatenation, simply put, is the process of linking the data vectors from different sources together to form a larger input vector. (4).

[0048] In one embodiment, the prediction time window is obtained through a load prediction model. Total demand Total available transport capacity and load rate The specific details are as follows: Define time window The time interval covered is: , Indicates the duration of the window, from which the window begins to open. Beginning, to Finish; Define time window Available transport capacity for: (5) in, For time window Number of departures This refers to the number of seats per vehicle. The duration of the time window (in hours); Time window based on machine learning model Demand forecast Defined as: (6) Where A is the confidence feature vector output in step 1. Describe the time window Environment state; function It is a trained prediction model that has the ability to process multidimensional nonlinear relationships, heterogeneous features and time series data. The model can be selected from gradient boosting tree model (XGBoost), random forest, multilayer perceptron, deep neural network, long short-term memory network or gated recurrent unit. Time window The load rate is: (7) The customer segmentation and willingness prediction module includes a customer segmentation module and a willingness prediction module; Customer segmentation module, based on ticket-purchasing passengers eigenvectors Input into a lightweight wide-depth mixture prediction model, ticket-buying passengers eigenvectors for: (8) in, This refers to user profiles, with typical fields including age, place of residence, travel frequency, and average advance ticket purchase time. Refers to the time window Environmental changes, including hourly characteristics, holidays, weather, flight peaks, and flight density; Time window Internal load rate; This refers to the compatibility between a user's historical travel routes and their target route. Refers to the most recent n ticket purchase records and payment channels; The probability of the passenger making an immediate purchase under a given price condition, output by the wide-depth hybrid prediction model, is denoted as: (9) in, Indicates passenger In the time window The event of completing the ticket purchase within the specified time; It is a candidate discount or offer level in the system; F𝑢( Then inform the model of this tourist. Who is it, what is the current situation, and what price are we offering? Based on real-time purchase probability Passengers The system automatically categorizes customers into three types: existing customers, wavering customers, and low-interest customers; and provides services for travelers. In the time window Assign a discrete label ; The willingness prediction module makes predictions based on customer categories: Existing customers still have a high probability of making a purchase even without discounts: , Taking experience as an example, it is believed that passengers They will generally buy unconditionally, so discounts have little impact on them; their existing customers are those with essential needs. Swing customers are less likely to buy when there is no discount, but their purchase probability increases significantly when there is a moderate discount: , in Discounts are the threshold for driving volume increases. Swing customers are price-sensitive customers; Low-intention customers: These are customers who, even with significant discounts, are still less likely to purchase than a certain threshold. Low-willingness customers are weakly related customers.

[0049] In one embodiment, the dynamic pricing and demand processing module further includes an environmental status input unit, an action selection unit, an execution and feedback unit, a reward and revenue unit, and a strategy update unit; The environment status input unit is used to input status information including system load rate, customer profile vector, and time information; system load rate Based on the predicted demand With total available capacity Calculations show that The target passenger group profile aggregation vector is generated by statistically analyzing the characteristics of all passengers using an aggregation function. Time characteristics include hourly features, holidays, and flight peaks; The system also utilizes the load rate to construct a discount cap constraint function (i.e., a Sigmoid smooth discount adjustment function): (10) in, Indicates the maximum allowed discount; For the Sigmoid function; Indicates adjustment of steepness; This represents the center point of the Sigmoid function; the upper limit of the discount is used as "part of the environment state" to constrain the agent's action set.

[0050] The action selection unit is the agent's choice of a discount action from a predefined discount set {0%, 10%, 20%, 30%}. ; The selected action must meet the following requirements: For candidate discounts, the discount must be within the allowed range (i.e., the feasible discount range). The corresponding bus fare is calculated as follows: ,in, This is the base fare.

[0051] The execution and feedback unit is where the system quotes a price based on the selected discount, obtains the user's purchase result and actual revenue; the probability of a single passenger purchasing is... During the time window The average purchase probability for domestic travelers is: (11) The expected number of actual transactions is: (12) Discount The expected return is: (13) in, This is a cost function related to discounts. Discount-related costs include discount concessions and channel promotion expenses, defined as: , This is the discount cost coefficient; the actual and expected returns will be entered into the reward unit. The reward unit calculates the reward value based on the difference between the actual reward and the maximum reward: ,in, This represents the baseline expected return with or without a discount; reward. Used to measure actions The advantages and disadvantages; The policy update unit is where the agent updates its policy parameters using Q-learning or DQN algorithms to maximize long-term returns. In the reinforcement learning DQN framework, the system finds the optimal discount level by searching or parsing an approximation within an allowed interval.

[0052] (14) Simultaneously satisfying the constraints

[0053] in, The minimum allowed ticket price, The maximum magnitude of the two price adjustments, For the previous time window The discount, or allowed range, refers to a feasible discount domain jointly defined by the constraints. The revenue sharing and operational coordination module includes a revenue sharing unit and an operational coordination unit; Revenue sharing unit, used for real-time calculation and attribution to all parties; settlement period is a time window. The intermodal transport transactions actually completed within the region constitute a distributable revenue pool, and the distributable net revenue of the revenue pool is... Let the set of partners who participated in this intermodal transport service and contributed to the transaction be: The value function of the participants is defined as follows: ,in, It is a subset of any set of participants. This indicates if only this group of participants is collaborating; the classic Shapley value represents each participant. Define the weighted average of its marginal contributions: (15) When the The expected value of a party joining any existing partner S is calculated, and the expected value is calculated for all possible joining orders. The order in which all participants join the system is random, and the contribution of the same participant will change depending on the joining order. Therefore, the expected value is calculated for all joining orders, i.e., the average value is calculated. A two-step pruning approximation Shapley value algorithm is used as both the contribution pruning algorithm and the alliance value pruning algorithm. The algorithm content is as follows: The system employs contribution-based pruning, first estimating the upper bound of the marginal contribution of each participant. : (16) in, This represents the potential revenue generated when all stakeholders in the system participate. This represents the revenue that the system can generate after removing subject i.

[0054] Then the participating parties are considered to be within the time window. The marginal effect of those involved is minimal and they are not included in the real-time amortization calculation set, resulting in a reduced set of core participants: (17) in, The contribution threshold; In the core stakeholder group For the remaining few core participants, the system uses Monte Carlo sampling to estimate the Shapley value, obtaining an approximate Shapley contribution for each retainer, denoted as . Ultimately, each retainer was obtained. Weighting allocation: (18) According to the time window Distributable net income of the income pool Divide proportionally: (19) in, The participants are within the time window The corresponding settlement amount; The operational coordination unit is used by the system to plan for future revenue sharing. The load rate prediction is calculated for each consecutive time window:

[0055] Set a capacity strain threshold When satisfied , It is a natural number, meaning the system's predicted value remains under high load; the system generates capacity intervention suggestions, defined as a suggestion vector: (20) in, For the suggested increase in the number of classes, For the suggested increase in seating capacity, High-risk routes are numbered. The specific time window to be added as suggested; When the system predicts that the load rate exceeds the threshold At that time, the system generates a capacity suggestion vector and sends an early warning to the operations dispatch center. This specific implementation provides a system for dynamic processing of intermodal ticketing based on load forecasting and implicit customer data. It addresses the technical obstacles caused by reliance on a single, delayed, or erroneous external data source, leading to inaccurate pricing decisions and even user complaints. It overcomes the shortcomings of traditional fixed-discount intermodal tickets, which cannot adapt to real-time market supply and demand changes, resulting in lost revenue during peak periods and idle seats during off-peak periods. It also solves the problem of traditional discount strategies failing to differentiate customers, wasting significant marketing resources on "existing customers" who will consume regardless. Furthermore, it addresses the need for quantified inventory data in dynamic pricing algorithms, as the crude revenue distribution mechanism in intermodal cooperation makes it difficult to fairly measure the contributions of all parties. This invention introduces a time-window system load rate model, remapping any valid continuous pass scenario for any train into a series of calculable expected load data within discrete time windows. A measurable digital twin is established, enabling real-time calling and optimization by the dynamic pricing engine. The system combines macro-level load forecasting with user intention modeling to achieve forward-looking perception of market demand and adaptive price generation, thereby realizing dynamic adjustment of fares, precise guidance of demand, and intelligent allocation of revenue. This truly endows the intermodal ticketing system with intelligent revenue management capabilities.

[0056] Example 1 This invention provides a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customer relationships, such as... Figure 2 As shown, Step 1: Data Acquisition and Fusion.

[0057] Responsible for multi-source data access, cleaning, and confidence calculation. Inputs include flight schedules, bus departure information, sales records, weather and traffic data, and outputs a standardized data stream with confidence levels.

[0058] For input flight data and other external information, a trust scoring model is used: 【 = exp(-(target time – timestamp)); = Data comparison with historical baseline and "lightweight statistical tests" (t-test and approximate KS test) complete data quality identification and repair within 5 milliseconds, ensuring that the source data input to the prediction model is clean and reliable.

[0059] The system in each time window (e.g., time window) Within the system, raw data is acquired in real time from multiple external business sources. Let these data sources be: The source of a flight (flight schedule / flight arrival forecast) is denoted as source.

[0060] The source of bus dispatch (bus schedules and vehicle availability) is denoted as source.

[0061] The source of ticket sales (real-time ticket sales records) is recorded as the source.

[0062] External environmental factors (weather, road conditions, holidays, etc.) are denoted as sources.

[0063] For any data source i, the system assigns that source a time window The collected set of key indicator vectors is denoted as ,For example: Flight arrival number, scheduled arrival time, and delay status. Number of tickets sold, average transaction price Planned departure times and number of seats available per vehicle Weather status code, traffic congestion index The system will not directly... Instead of feeding the data into the subsequent model, the system first determines whether the source data is "credible" within the time window. To this end, the system calculates a trust score (credibility rating) for each source i. Defined as: (1) in The "timeliness" of the data can be defined as follows: , Indicated as source The latest data update time is far from the time window Time difference (minutes) The time decay constant, Represented as a time window The first Data sources. For example, = 10 means that a 10-minute delay will significantly reduce freshness.

[0064] This indicates that the source data is consistent with the system reference baseline.

[0065] Defined as: (2) in, It can be a historical average, a weighted average of the same indicator from other sources, or a smoothed value from the previous time window; To prevent tiny constants with a denominator of 0. The closer to 1, the smaller the deviation and the more reliable the data.

[0066] , For the weighting coefficients, satisfying + =1, used to balance "timeliness priority" versus "stability priority". In a typical implementation, this is optional. =0.6, =0.4.

[0067] Then, the system uses a global threshold. (e.g., 0.75) Examine the data quality for each source: like If so, the data from that source is considered reliable; like If the data is unreliable, it is considered low-reliability data and needs to be corrected.

[0068] For low-reliability data, the system will... Perform a quick correction. The correction method can be a weighted average with historical smoothed values, for example: (3) in, This source was in the previous time window. Or a smoothed reference value for the same time period on the previous day; It can be set according to business importance and scenario reliability. After correction Only then will it be adopted.

[0069] Finally, the system stitches together reliable (or corrected) data from all sources to obtain the time window. Confidence feature vector: (4) this This is the standardized confidence level data input that is directly used in the subsequent "load prediction module (step 2)" and "customer intention prediction module (step 3)".

[0070] The following uses a time window (e.g., 08:00–09:00) to demonstrate a complete round of credibility calculation and fusion process.

[0071] Assuming the system is within the time window The following information was collected: Flight source : 22 flights are expected to arrive this hour. (Last updated time since...) = 2 minutes, compared with reference data (previous time period and other sources), the flight volume deviation is very small. ≈ 0.95 Sales Source 180 connecting tickets have been sold in the last hour. (Last updated) = 1 minute, the difference between sales volume and the previous window / historical average is acceptable. ≈ 0.98 Bus dispatch source : 10 departures are expected this hour, each with 50 seats (theoretical capacity 500 seats). Latest update time. = 8 minutes. Due to factors such as last-minute vehicle changes and delayed departures, the historical comparison may deviate slightly. ≈ 0.90 External environmental sources Weather: Light rain (code 1), City traffic congestion index 0.62, Last updated time = 6 minutes, with moderate deviation compared to historical data. ≈ 0.92 set up =0.6, =0.4, =10 (time constant 10 minutes), calculate for each source : Regarding flight data:

[0072] Regarding sales data:

[0073] Bus dispatch data:

[0074] External environment data:

[0075] Now compare the credibility score of each source with the threshold. =0.75: Flight data The data is reliable and can be used directly.

[0076] Sales data The data is reliable and can be used directly.

[0077] Bus dispatch data The data is of low reliability and needs to be corrected. External environment data The data is of low reliability and needs to be corrected. For sources requiring correction (bus scheduling and external environment), the system performs weighted smoothing correction. For example, for the indicator "10 planned departures" in bus scheduling, if the smoothing reference value for the same time period and adjacent time windows on the previous day is 11 departures, then it can be corrected by... Calculate the correction value:

[0078] The system can round it to "10 flights", but will internally mark it as "low confidence in scheduling data" so that subsequent prediction models will be cautious when dealing with capacity (e.g., by reducing the available capacity assumption through a reduction factor).

[0079] Similarly, external environmental data (such as traffic congestion index) can also undergo the same smoothing process to prevent occasional abnormal peaks from being directly amplified in subsequent models.

[0080] After the above credibility screening and correction, the system forms a time window. Confidence feature vector For example, in this example, It can include the following standardized fields (illustrated): Flight arrivals (forecast for this window): 22 flights Expected Passenger Inflow Intensity Characteristics: Derived from Flight Arrivals and Load Factor Ticket sales in the last hour: 180 tickets Bus plan effective capacity: approximately 10 trips × 50 seats / trip = 500 seats (with low credibility rating) Weather code: 1 (light rain) Urban congestion index (smoothed): 0.62 Holidays: 0 Competitive transportation accessibility: Comparative features such as "High-speed rail fare for the same section = ¥45, average taxi fare = ¥110". Confidence feature vector. It will be directly fed into the load forecasting model in step 2 to calculate the demand forecast for that time window. Available transport capacity and load rate .

[0081] Step 2: Calculation of system load rate within the time window Predicting the time window of scrolling events using machine learning models The system load rate (i.e., the ratio of demand to capacity) is calculated by inputting the fused data from step 1 and outputting a real-time load rate sequence.

[0082] 1) Time window Segmentation: The system segments the operating period into fixed-length time windows.

[0083] Define time window Divide the operating time into continuous time windows. (For example, using 1 hour as the unit), let the window length be... Hours (e.g.) =1); for any time window We define the time interval covered by this window as follows: The system in each time window Estimate the available capacity and potential demand for the next hour, and calculate the forecast load factor for that window. .

[0084] 2) Window capacity estimation: Calculate the total capacity within the window: Calculate the capacity for each time window based on bus departure frequency and the number of seats per bus. The total number of seats that the internal bus system can provide (total capacity = frequency of service × number of seats per vehicle × window duration).

[0085] For time windows The system first calculates the time window. Available transport capacity within This value represents the theoretical maximum number of passengers that the intermodal bus system can transport within this time window. The definition is as follows: (5), in, For time window Number of departures This refers to the number of seats per vehicle. The duration of the time window (in hours).

[0086] 3) Demand window forecasting: Predicting total demand within the window: Using machine learning models (such as XGBoost) combined with data such as flight schedules, historical passenger flow, and the number of tickets already sold, predict the total demand within that time window. Total passenger demand.

[0087] The system uses a machine learning model to analyze the time window. Predicting travel demand , Defined as: (6) Where A is the confidence feature vector output in step 1. Describe the time window Environment state; function It is a trained prediction model capable of handling multidimensional nonlinear relationships, heterogeneous features, and time-series data. The model can be a gradient boosting tree model (XGBoost) or other machine learning models. The input features of the model may include, but are not limited to, the expected number of flights arriving in the corresponding time period of the window, the real-time ticket sales speed of the previous time period, the target price / discount level, weather conditions, road condition index, holiday signs, passenger flow patterns of the same historical time period, and the availability and cost level of competing transportation modes (high-speed rail / ride-hailing).

[0088] Model output This represents the number of potential passengers expected to purchase or use the intercity bus service within this hour (or the time period defined by the window).

[0089] 4) Calculate the system load factor: System load factor = Total demand within the forecast window / Total capacity within the window. The real-time changing load factor will replace the micro-level "remaining seats" and become the core decision-making basis for the entire pricing system.

[0090] After obtaining available transportation capacity and forecasting demand The system then calculates the load rate for that window.

[0091] Time window The load rate is: (7).

[0092] This value can be understood as the degree of full load or resource scarcity, and is used to determine whether the target period is characterized by oversupply (i.e., low load rate) or undersupply (i.e., high load rate).

[0093] The following uses a set of specific data to demonstrate the load rate calculation process; The scheduling system in the time window Internal plan =10 trips, each with 50 seats. Assuming that a small number of vehicles may experience delays or have uncertain attendance, the system introduces an average availability coefficient of 0.95 (meaning that on average only 95% of seats are actually available per trip). Therefore, the effective seating capacity can be written as:

[0094] In the same time window The system will use the confidence characteristics of this time period. (For example: estimated number of arriving flights = 22, sales in the past hour = 180 orders, weather = light rain, road condition index = 0.62, whether it is a holiday = no, target average ticket price range = 36~40 yuan, etc.) Input into the trained prediction model Assuming the model outputs: =440, which means that the system expects approximately 440 passengers to have actual ticket purchase / ride demand for the "Airport-City Bus Intermodal Service" during the period of 08:00–09:00.

[0095] Substitute the two results above into the formula: =0.926, this result will be used as input to pricing. The module (step 4) triggers high-load strategies (such as limiting large discounts and increasing price elasticity to protect peak revenue), which are synchronously transmitted to the scheduling and coordination module to determine whether it is necessary to add temporary shifts or increase vehicle seating configuration.

[0096] Step 3: Customer intention prediction and implicit segmentation.

[0097] This study uses a hybrid wide-depth forecasting model to analyze individual purchase intentions and price sensitivity, determining whether individuals are price-sensitive or have essential needs, and calculating purchase intentions.

[0098] Establish a system load forecasting model (XGBoost): Its task is to perform calculations on the "total demand within the forecast window" and output the macroscopic system load rate.

[0099] Establish a purchase intention prediction model (wide & deep): Its task is implicit customer segmentation. It uses multi-dimensional data such as user historical behavior and travel characteristics to accurately predict the target user's sensitivity to different discounts, i.e., the probability of purchase. This effectively distinguishes between "existing customers" and "swing customers".

[0100] For time windows One of the candidate ticket buyers The system constructs a set of feature vectors for the passenger, denoted as... This feature vector may include, but is not limited to: individual passenger profile features (such as past ticket purchase frequency, travel scenario attributes, whether tickets were purchased close to the departure time, and historical responsiveness to discounts), and target sales features (system load rate within the given time window). External conditions such as average base fare within the window, channel type, weather, and transportation accessibility, as well as behavioral proximity characteristics (such as the degree of matching with flight arrival time and destination and route coverage), can be uniformly represented as: (8) The system will include the above features The input is fed into a lightweight wide-depth mixture prediction model. The model's role is to output the instantaneous probability of the traveler making a purchase given a price. This probability is denoted as: (9) Indicates passenger In the time window The event of completing the ticket purchase within the specified time; It refers to a candidate discount or offer (e.g., a 10% reduction) considered by the system; quantity Then tell the model "who this person is, what the current situation is, and what price we are offering".

[0101] Yes The system can then infer the passenger's price sensitivity and demand rigidity, automatically classifying the passenger into different customer types to guide subsequent pricing. Optional segmentation logic is as follows (example only; generally, a fixed threshold is used after training): Existing customers (those with essential needs): If there are no discounts (e.g.) Even with a 0% chance of failure, there is still a relatively high probability of purchasing. This suggests that the passenger would likely make the purchase unconditionally, and the discount would have little impact on them. An empirical value of 0.7 or 0.8 can be used.

[0102] Swing customers (price-sensitive): They are less likely to buy when there is no discount, but their probability increases significantly when there is a moderate discount (e.g., D=10%). ,in This is the threshold for "discount-driven volume increase," for example, 0.2 (meaning an increase of more than 20 percentage points). These people are the best target for discounts and belong to the core group for "mining incremental growth."

[0103] Low-interest customers (weakly related): those who are offered higher discounts (e.g.) (20%), the probability of purchase is still below a certain lower limit: , A rate of 0.3 is acceptable. These types of passengers are not attractive enough for this intermodal transport product and are unlikely to be converted in the short term.

[0104] Based on the above rules, the system will provide passengers with... In the time window Assign a discrete label These tags will then directly participate in step 4 (dynamic pricing and demand management) to determine whether and to whom discounts are offered.

[0105] The following constructs a time window = A specific example of 08:00–09:00, assuming we are analyzing a passenger (Just arrived at the airport, preparing to purchase a "flight + bus" combined transport segment) Target window load rate =0.926 (from step 2), indicating that resources were scarce during this period; Passenger profile characteristics: Purchased tickets for this route 3 times in the past 30 days; arrival time is close to the morning rush hour; past purchase records show frequent ticket purchases even without discounts; among alternative travel options, ride-hailing costs approximately ¥110, significantly higher than the bus. Encoding this information forms... .

[0106] The system attempts to test several discount candidate levels for the passenger. And calculate the purchase probability for each: No discount ( =0%), That is, even without a discount, the probability of him placing an order is 78%.

[0107] Medium discount ( =10%, equivalent to a slight reduction in ticket price. The discount only resulted in a 5 percentage point increase.

[0108] Larger discounts =20%) Larger discounts further increase the probability of purchase, but the increase is only 2 percentage points.

[0109] Based on the threshold values ​​set earlier: =0.7, then This indicates that the passenger is still highly likely to make a purchase even with zero discount, meeting the criteria for "existing customer (essential customer)".

[0110] In contrast to another type of passenger (Assuming late arrival and price sensitivity), the calculation yields... They were identified as "swing customers (price-sensitive)". In other words, these people are the main target of discount campaigns: giving them a discount can significantly increase sales, thereby improving seat utilization and reducing idle time during low-load periods.

[0111] In this step, the system targets the time window. There will be two types of core outputs: 1. A "discount response curve library", that is, for different discount candidates Purchase probability estimation for different user types: Essentially, it's about "If I lower the ticket price to this level, how many people will actually buy it?"

[0112] 2. A set of user segmentation tags: The subsequent pricing module will then determine which users would buy even without a discount (no additional concessions needed); which users would be converted with just a slight price reduction (discounts should be focused on them); and which users wouldn't be attracted even with price reductions (no excessive financial sacrifices required).

[0113] Step 4: Dynamic Pricing and Demand Management Dynamic pricing and demand management involves calculating the optimal discount decision and outputting a dynamic discount. The objective function is passed to the front-end ticketing system via an interface. This system acts as the decision-making brain, its objective no longer being simple pricing, but proactive demand management. The decision-making logic is primarily based on the "system load rate" and the "customer type" derived from the willingness model.

[0114] When the load factor is low (<50%): With ample capacity, the system will offer more attractive discounts to "swing customers" to boost off-peak revenue.

[0115] When the load factor is high (>80%): Due to capacity constraints, the system will automatically reduce or even cancel discounts, offering only full-price or near-full-price connecting tickets to all customers (including those on a tight schedule). This is a proactive "price increase" strategy to smooth out excessive demand during peak periods and ensure service quality.

[0116] Objective function: In any decision, the system aims to maximize the expected net profit, ensuring that all discounts are issued on the premise of creating incremental revenue.

[0117] The technical solution of this invention lies in constructing and solving a unified profit maximization objective function, integrating macro-level supply and demand forecasting with micro-level passenger willingness modeling into the same optimization framework, and realizing minute-level automatic ticket price decision-making.

[0118] The specific tasks of this step are as follows: Figure 3 As shown, this is based on the system's real-time load rate. Customer purchase probability distribution It automatically calculates the optimal ticket price or discount level for a time window. To achieve maximum revenue, optimized capacity, and a balanced experience.

[0119] (1) Environmental status input In the time window The system reads all environment variables generated in steps 2 and 3 and constructs a time window. The state vector includes, but is not limited to: system load rate, confidence feature vector, passenger purchase intention, passenger segmentation structure, etc.

[0120] The system is based on the load rate The target window is divided into three states, as shown in Table 1: Table 1

[0121] To prevent prices from jumping at critical points, the system uses a sigmoid smoothing discount adjustment function: (10) in, Indicates the maximum allowed discount; For the Sigmoid function; This indicates the adjustment of steepness (typically 10–20). This represents the center point of the Sigmoid function (typically 0.65–0.7).

[0122] The following example calculation uses a set of data. Assume a time window... The status from 08:00 to 09:00 is as follows, see Table 2: Table 2

[0123] The maximum allowable discount is now calculated based on the load factor: =0.5%. This means that under high load conditions, the system can only allow a tiny discount of about 0.5% (i.e., almost no price reduction), which is consistent with the logic of a "supply shortage" scenario.

[0124] (2) Action selection The system selects a discount action based on the target state. The selected action must meet the following requirements. That is, it must be within the allowed range (feasible discount range) for candidate discounts. The corresponding bus fare is calculated as follows: ,in, The base fare is (e.g., ¥40).

[0125] Assume the system allows searching. The discounts and corresponding ticket prices are as follows, see Table 3: Table 3

[0126] Let the predicted demand within this time window be Available transport capacity is The upper limit of potential ticket purchasers For any candidate discount D, the corresponding ticket price is: ,in This is the base fare.

[0127] (3) Execution and feedback The system quotes a price based on the selected discount and obtains the user's purchase results and actual revenue.

[0128] If within the time window The average purchase probability of the target passenger group is: (11) The expected number of actual transactions is: (12) Therefore, the expected return under this discount is: (13) in The cost function related to discounts (including discount concessions and channel promotion expenses) can be defined as: , This is the discount cost factor (e.g., a loss of 0.01 units of revenue or promotional expenses for every 1% discount).

[0129] Assume the system still allows searching. We input our respective purchase probabilities and costs to calculate the expected returns, as shown in Table 4: Table 4

[0130] (4) Rewards The system is based on the time window The difference between actual and theoretically optimal returns is used to construct the reward value, which is then used for policy learning. (Reward income) ,in This represents the baseline expected return under the reference discount (or no discount). Reward Used to measure action The advantages and disadvantages.

[0131] Assume the system still allows searching. The reward benefits under different discounts are as follows, see Table 5.

[0132] Table 5

[0133] (5) Strategy update The agent updates policy parameters through Q-learning or DQN algorithms to maximize long-term returns.

[0134] The system finds the discount level that maximizes the profit by searching (or analytically approximating) within the allowed interval: (14) Simultaneously satisfying the constraints

[0135] The minimum allowed ticket price, The maximum range of two price adjustments (to prevent frequent fluctuations).

[0136] The reward process results show that the optimal discount is around 10% under low load. However, at this point (Load=0.926), the discount cap is constrained to 0.5% by the Sigmoid barrier, therefore the final discount implemented by the system is: This means that only a symbolic discount of about 0.5% is offered (or the original price is maintained).

[0137] The system will recalculate every few minutes. , and If the future time window Predicting load drop (e.g.) If the discount is 0.45, the discount will be automatically increased to: =28.6% At this point, the system can offer a discount of approximately 28% to stimulate the conversion of "wavering customers".

[0138] Step 5: Revenue sharing and operational collaboration.

[0139] The revenue contribution of each partner is calculated based on the two-step pruning approximation Shapley value algorithm, realizing the linkage between the revenue algorithm and capacity scheduling.

[0140] Revenue Sharing: After a successful transaction, an approximate Shapley value algorithm based on two-step pruning (contribution pruning and alliance value pruning) is used to calculate and attribute the revenue to each party in real time and fairly, and record it in the virtual account.

[0141] Capacity scheduling and operational coordination: The system continuously monitors flight status and can automatically handle services such as extending associated bus tickets. Simultaneously, when it predicts that the system load will be excessively high, it can automatically send an alert to the operations dispatch center, suggesting the addition of more flights.

[0142] Step 5 involves two tasks. The first task is to calculate the revenue contribution share of each participant (such as airlines, bus companies, channel platforms, dynamic pricing system modules, etc.) in a single intermodal sales transaction or time window, and to distribute the profits accordingly. The second task is to transform high load risk (i.e. capacity shortage) into capacity scheduling suggestions, so as to form a closed loop between pricing strategies and resource allocation.

[0143] 1) Calculation of revenue sharing Let a settlement period of this system be a time window. (For example, the hour 08:00–09:00). The intermodal transactions actually completed within this window constitute an allocable "revenue pool." We denote the allocable net revenue of this revenue pool as... Let the set of partners who participated in this intermodal transport service and contributed to the transaction be: For example: A. Airline (provides passenger flow entry), B. Bus operator (provides actual seats), C. Sales channel / platform (such as OTA channel, bringing exposure and conversion). Dynamic pricing module (the pricing decision engine in this invention, which improves conversion and revenue).

[0144] We introduce a value function for these participants. ,in It is a subset of any set of participants. This indicates the potential benefit if only this group of participants collaborated (with the others not participating or working at the baseline level). This forms the basis for Shapley value amortization.

[0145] Classic Shapley values ​​for each participant Define the weighted average of its marginal contributions: (15) When the What is the increase in revenue when Fang joins any existing partner S? And what is the expected value for all possible joining orders?

[0146] To achieve real-time performance, this invention employs a two-step pruning approximation Shapley value algorithm.

[0147] 2) Two-step pruning approximation Shapley value algorithm Step 1: Contribution Filtering and Pruning. The system first estimates the upper bound of the marginal contribution of each participant. : (16) ( If a participant's contribution threshold is set (e.g., 1%-2% of total revenue), then that participant is considered to have minimal marginal impact on the window and is temporarily excluded from the real-time allocation calculation set. This results in a reduced set of core participants. : (17) in, The contribution threshold; Step 2: In For the remaining few core participants, the system uses Monte Carlo sampling to estimate the Shapley value, rather than performing an exact traversal of all subsets. This yields an approximate Shapley contribution for each retainer, denoted as [missing value]. Ultimately, each retainer was obtained. Allocation weights : (18) And based on this, the window revenue pool Divide proportionally

[0148] , Is the participating party at the window The corresponding settlement amount.

[0149] Continue using the previous scenario of running the window from 08:00 to 09:00.

[0150] Assuming the window eventually implements a dynamic pricing strategy. Subsequently, the actual transaction generated distributable net income. =14145 The participating parties are assumed to be: Airline A, Bus Company B, Online Travel Agency (OTA) C, and Dynamic Pricing Engine D (the decision-making module of this invention). The system obtains the following approximate upper bound of marginal contribution through backpropagation or A / B comparison:

[0151] Assuming a threshold =300, for all participants ≥300, therefore The system then The Shapley value is obtained using the Monte Carlo approximation. For example, we get:

[0152] Normalize it to obtain the weights

[0153] Finally, the time window Profit Pool =14145 is allocated: airlines Bus ;channel Pricing engine

[0154] After revenue sharing, it is also necessary to coordinate transportation resources and the system for future... The load rate prediction is calculated for each consecutive time window: ; Calculate the load factor forecast (e.g., for two or three rolling time windows in the future). Set a capacity shortage threshold. (e.g., 0.85), when satisfied This means that the system predicts that it will continue to be under high load, rather than experiencing occasional spikes.

[0155] Once a sustained high load risk is triggered, the system generates capacity intervention suggestions. This can be defined as a suggestion vector: (20) in The suggested number of additional classes (e.g., +1 class). For the suggested seat increment High-risk routes are numbered. The specific time window to be added (e.g., "09:00–10:00").

[0156] This embodiment provides a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers. It addresses the technical obstacles that lead to inaccurate pricing decisions and even user complaints due to reliance on a single, delayed, or erroneous external data source; overcomes the shortcomings of traditional fixed-discount intermodal tickets, which cannot adapt to real-time market supply and demand changes, resulting in lost revenue during peak periods and idle seats during off-peak periods; resolves the fundamental business pain point that traditional discount strategies cannot differentiate customers, wasting significant marketing resources on "existing customers" who will consume regardless, thus eroding net profit; addresses the fundamental contradiction between the "quantified inventory" data required for dynamic pricing algorithms and the actual operating model of bus tickets being sold "not by seat, or even not by trip (valid for any trip)"; and resolves the business problem in intermodal cooperation where a crude revenue distribution mechanism makes it difficult to fairly measure the contributions of all parties, hindering long-term stable cooperation. This invention introduces a "time window system load rate" model, remapping the continuous pass scenario of "valid for any trip" into a series of calculable expected load data within discrete time windows. In this way, a measurable digital twin is established for the fuzzy pass inventory, enabling it to be called and optimized by the dynamic pricing engine in real time. Furthermore, the system combines macro-load forecasting with user intention modeling to achieve forward-looking perception of market demand and adaptive price generation, thereby realizing dynamic adjustment of ticket prices, precise guidance of demand, and intelligent allocation of revenue; fundamentally solving the long-standing technical bottleneck of "uncalculable inventory under the through-ticket system", enabling the intermodal ticketing system to truly possess intelligent revenue management capabilities.

[0157] Example 2 In one embodiment, the load forecasting model (step 2 in the method) further includes the following scheme.

[0158] Any machine learning model capable of effectively processing structured data and performing regression predictions can be used. For example, ensemble learning models such as Gradient Boosting Decision Tree (GBDT) and Random Forest can be employed. For passenger flow with significant time dependencies, deep learning time series models such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GR) can also be used.

[0159] Feature expansion: The input features used for prediction, in addition to the features mentioned in the preferred embodiments, may further include, but are not limited to, a combination of one or more of the following: real-time weather data (such as rainstorm and snowstorm warnings), real-time security check queue times at the origin airport, urban traffic congestion index, whether a large conference or cultural and sports event is being held locally, and ticket prices or availability data of competing modes of transportation (such as high-speed rail and taxis).

[0160] This invention provides a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customer data. The load forecasting model using the aforementioned alternative has the following advantages: First, it possesses stronger model generalization capabilities. By supporting multiple machine learning or deep learning models, the system can flexibly select the optimal forecasting model based on different routes, seasons, or data distributions, improving the accuracy and stability of load forecasting. Second, it offers higher-dimensional scenario adaptability. By introducing extended features such as weather, traffic, and events, the model can identify more external factors affecting passenger demand, enabling it to cope with special situations and emergencies. Third, it enhances operational control capabilities. More accurate demand forecasting improves the decision-making quality of discount optimization and capacity scheduling, allowing the intermodal system to identify demand peaks in advance and achieve closed-loop management of "prediction-intervention-scheduling."

[0161] In one embodiment, the intention prediction model (step 3 in the method) also includes the following technical solutions. The core of the intention prediction model in this invention is predicting the conversion probability (CTR) of users in a specific scenario. Therefore, other models widely used in this field can also be applied. For example, in scenarios where model simplicity and interpretability are desired, a logistic regression model can be used; to enhance feature cross-cutting capabilities, a factorization machine (FM) or its variants (such as FFM) can be used; and in scenarios where higher prediction accuracy is desired, other deep learning models, such as DeepFM and xDeepFM, can also be used.

[0162] In one embodiment, the dynamic pricing strategy (step 4) also includes the following two technical solutions. The continuous function in the dynamic pricing strategy of this invention: discount level It can be modeled as system load rate A continuous function, namely = .For example, It can be a linearly decreasing function or the inverse of a smoother sigmoid function, so that the higher the load factor, the smoother the discount decreases.

[0163] The intention prediction model employing the above-mentioned continuous discount alternative of the present invention has the following beneficial effects: First, it improves the accuracy of discount adjustment, extending the discount from discrete action to continuous adjustment, which helps to control revenue with finer granularity; it can avoid excessive jumps such as 10%→20%; Second, it improves the smoothness of dynamic pricing and user experience, as the continuous discount function can ensure that discount changes are natural, reduce passengers' sensitivity to price fluctuations, and improve overall user satisfaction.

[0164] This invention relates to a reinforcement learning DQN framework for dynamic pricing strategies: in a more advanced implementation, pricing decisions can be made by a reinforcement learning agent. Within this framework: State: Includes the target's system load rate, user profile characteristics, time information, etc.

[0165] Action: Select one from a discrete set of discounts (such as {0%, 10%, 20%, 30%}) or output a continuous discount value.

[0166] Reward: When a user completes or abandons a purchase, the optimization objective function E[ according to the present invention] is applied. The agent is rewarded positively or negatively based on the actual value achieved, thereby learning the optimal dynamic pricing strategy through continuous interaction with the environment.

[0167] This invention provides a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customer behavior, offering advantages over the aforementioned alternatives. Compared to traditional static or rule-based pricing methods, the reinforcement learning pricing framework of this invention offers the following advantages: First, it can adaptively learn the optimal pricing strategy without manually setting complex rules; the system automatically discovers the optimal discount pattern through interactive learning. Second, it can capture user behavior feedback; Reinforcement Learning directly utilizes users' actual purchasing behavior through a reward function, which is closer to the real-world environment than prediction-based optimization. Third, it can adapt to nonlinear and time-varying market scenarios; Reinforcement Learning can handle complex nonlinear relationships, such as sudden surges in passenger flow caused by holidays, severe weather, or unexpected events.

[0168] In one embodiment, the revenue sharing mechanism (step 5 in the method) also includes the following technical solution. In the revenue sharing mechanism of this invention, any game theory algorithm or contribution evaluation model aimed at fairly distributing cooperative revenue can be applied. For example, the classic Amann-Shapley algorithm can be used. In scenarios with extremely high requirements for real-time computation, a simplified dynamic weighted model based on marginal contribution rate can also be used, that is, tracking the conversion rate of passenger traffic introduced by each partner (channel, airline) in real time and dynamically adjusting its revenue sharing weight accordingly.

[0169] This invention provides a method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers. The adoption of the aforementioned revenue-sharing alternative brings the following technical effects: First, it maintains the fairness and transparency of revenue distribution among partners. Both models distribute revenue from a "contribution" perspective, aligning with industry cooperation strategies and significantly enhancing partners' trust in the platform. Second, it is more suitable for complex intermodal scenarios. In multi-entity intermodal scenarios such as airlines, buses, and distribution channels, the dynamic contribution mechanism can reflect real-time changes in traffic generation value and sales capabilities, improving overall system revenue.

Claims

1. A method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers, characterized in that, The method steps are as follows: Step 1: Data Acquisition and Fusion In each time window Within this process, raw data is acquired, including data from K flights and real-time data from M external business sources; and data initialization is completed within a preset time to obtain a confidence feature vector. ; Step 2: Time Window System load rate calculation Confidence feature vector The data is fed into the load forecasting model to obtain the time window. Total demand Total available transport capacity And calculate the load rate. ; Step 3: Customer Intent Prediction and Implicit Segmentation Get Time Window Any candidate passenger eigenvectors ; will passengers Features Input into the wide-depth hybrid prediction model, the model outputs passenger data. Probability of immediate purchase given a price. ; Obtain passenger Price sensitivity and rigid demand, and will passengers Automatically categorize customers and assign them a discrete label. The subsequent pricing is guided by the segmentation logic; Step 4: Dynamic Pricing and Demand Processing Based on real-time system load rate With passengers Purchase probability distribution Automatic solution time window Best fare or discount level The discount must satisfy load constraints, regulatory constraints, and price smoothing constraints; Step 5: Revenue Sharing and Operational Collaboration Passengers After a transaction is successful, a two-step pruning approximate Shapley value algorithm is used to calculate and attribute the benefits to each party in real time. When the future time window The system load prediction meets the continuous load condition. At that time, the system generates a capacity suggestion vector and sends an early warning to the operations dispatch center.

2. The method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers according to claim 1, characterized in that, Step 1, data initialization, involves identifying and repairing the quality of the original data; specifically as follows: Step 1.1: Data initialization begins by determining if the source data falls within the time window. Whether it is reliable or not, the system is the data source. Calculate credibility score Defined as: （1） in, The timeliness of the data is defined as follows: , Indicated as source The latest data update time is far from the time window Time difference, The time decay constant, Represented as a time window The first Data sources; This indicates the consistency between each data source and the system reference baseline; it is defined as: (2) in, It is the historical average, the weighted average of the same indicator from other sources, or the smoothed value of the previous time window; Indicate source In the time window The original data values, To prevent tiny constants with a denominator of 0; , For the weighting coefficients, satisfying + =1; Step 1.2: Time Window The quality assessment and repair of the raw data is performed by the system using a global threshold. Examine the data quality from each source: like If so, the data from that source is considered reliable; like If the data is unreliable, it should be corrected. For low-reliability data, the system... Perform a fast correction; the correction method is a weighted average with historical smoothed values. （3） in, The data source is in the previous time window. Or a smoothed reference value for the same time period on the previous day; The value is set based on business importance and scenario reliability; Step 1.3: The system stitches together reliable or corrected data from all sources to obtain a time window. Confidence feature vector All sources refer to the system within the time window. The data sub-streams obtained from external modules include flight data, bus data, ticketing data, and weather data; Data concatenation involves linking the data vectors from different sources end to end to form a larger input vector. （4）。 3. The method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers according to claim 1, characterized in that, Time window in step 2 System load rate calculation involves using the confidence feature vector. Directly input into the load forecasting model to obtain the time window. Total demand Total available transport capacity and load rate , Define time window The time range covered is: , Indicates the duration of the window, from which the window begins to open. Beginning, to Finish; Define time window Available transport capacity for: （5） in, For time window Number of departures This refers to the number of seats per vehicle. The duration of the time window; Time window based on machine learning model Demand forecast Defined as: （6） Where A is the confidence feature vector output in step 1. Describe each time window Environment state; function It is a trained prediction model that has the ability to process multidimensional nonlinear relationships, heterogeneous features and time series data. The model is selected from gradient boosting tree model, random forest, multilayer perceptron, deep neural network, long short-term memory network or gated recurrent unit. Time window The load rate is: （7）。 4. The method for predicting and dynamically processing intermodal ticketing for implicit customers according to claim 1, characterized in that, Passengers purchasing tickets in step 3 eigenvectors for: （8） in, This refers to user profiles, with typical fields including age, place of residence, travel frequency, and average advance ticket purchase time. Refers to the time window Environmental changes, including holidays, weather, and flight density; Time window Internal load rate; This refers to the compatibility between a user's historical travel routes and their target route. Refers to the most recent n ticket purchase records and payment channels; eigenvectors The input is fed into a lightweight, wide-depth mixed prediction model; the model outputs the instantaneous purchase probability of the passenger given the price, denoted as: （9） in, Indicates passenger In the time window The event of completing the ticket purchase within the specified time; It refers to a specific candidate discount or offer within the system; Then inform the model of this tourist Who is it, what is the current situation, and what price was offered? Based on real-time purchase probability Passengers The system automatically categorizes customers into three types: existing customers, wavering customers, and low-interest customers; and provides services for travelers. In the time window Assign a discrete label ; Customers are categorized as follows: Existing customers: If they still have a high probability of making a purchase even without discounts: Then it is believed that the passenger He will basically buy unconditionally, so discounts have little impact on him. The value is taken from experience. Swing customers: They are less likely to buy when there is no discount, but their purchase probability increases significantly when there is a moderate discount. ; in, It is a threshold for discount-driven volume increase; Low-intention customers: those who, even with high discounts, are still less likely to purchase than a certain threshold. ; in , , Thresholds for classifying customer preferences; The threshold for increasing purchase volume driven by discounts.

5. The method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers according to claim 1, characterized in that, The system solves for the time window in step 4. Discounts within The discount ceiling is defined using the Sigmoid smooth discount adjustment function: （10） in, Indicates the maximum allowed discount; For the Sigmoid function; Indicates adjustment of steepness; Indicates the center point of the Sigmoid function; Let the predicted demand within the time window be... Available transport capacity is The upper limit of potential ticket purchasers For any candidate discount The corresponding bus fare is: ,in, The base fare; In the time window The average purchase probability for domestic travelers is: （11） The expected number of actual transactions is: （12） Discount The expected return is: （13） in, It is a cost function related to discounts, including discount concessions and channel promotion expenses; the cost function related to discounts is defined as: , This is the discount cost coefficient; The system finds the optimal discount level by searching within the allowed range. （14） Simultaneously satisfying the constraints in, The minimum allowed ticket price, The maximum magnitude of the two price adjustments, For the previous time window Discount; The allowable interval refers to a feasible discount region jointly defined by constraints.

6. The method for dynamic processing of intermodal ticketing based on load forecasting and implicit customers according to claim 1, characterized in that, The real-time calculation and attribution of benefits to each party in step 5 are as follows: Step 5.1: Settlement period is a time window The intermodal transport transactions actually completed within the region constitute a distributable revenue pool, and the distributable net revenue of the revenue pool is... ; Let the set of partners who participated in this intermodal transport service and contributed to the transaction be: The value function of the participants is defined as follows: ,in, It is a subset of any set of participants. Indicates only the participating subset Total revenue that can be obtained when participating in intermodal transport services; Step 5.2: Classic Shapley values ​​for each participant Define the weighted average of its marginal contributions: （15） When the Fang joins any existing group of partners What is the expected increase in revenue when adding the item, and what is the expected value for all possible order of addition? The order in which all participants join the system is random. The contribution of the same participant will change depending on the order in which they join. It is necessary to calculate the expectation of all joining orders, that is, to calculate the average value. Step 5.3: The two-step pruning approximation Shapley value algorithm consists of a contribution pruning algorithm and a federation value pruning algorithm, the details of which are as follows: Step 5.3.1: The system employs contribution-based pruning, first estimating the upper bound of the marginal contribution of each participant. : （16） in, This represents the potential revenue generated when all stakeholders in the system participate. Indicates the removal of the main body The benefits that the system can generate; Then the participating parties are considered to be within the time window. The marginal effect within the core participants is minimal and they are not included in the real-time amortization calculation set, resulting in a reduced set of core participants. : （17） in, The contribution threshold; Step 5.3.2: Gathering core stakeholders For the remaining few core participants, the system uses Monte Carlo sampling to estimate the Shapley value, obtaining an approximate Shapley contribution for each retainer, denoted as . Ultimately, each retainer was obtained. Allocation weights : ；（18） According to the time window Distributable net income of the income pool Divide proportionally : （19） in, It is the participants within the time window The corresponding settlement amount; Step 5.4: After revenue sharing, the system plans for the future. The load rate prediction is calculated for each consecutive time window: Set a capacity strain threshold When satisfied , It is a natural number, meaning the system predicts that the system will continue to operate at high load; the system generates capacity intervention suggestions. Define the suggestion vector as follows: （20） in, For the suggested increase in the number of classes, For the suggested increase in seating capacity, High-risk routes are numbered. The specific time window for the proposed addition.

7. A system for dynamic processing of intermodal ticketing based on load forecasting and implicit customers, characterized in that, The system includes a data acquisition and fusion module, a load forecasting module, a customer segmentation and willingness forecasting module, a dynamic pricing and demand processing module, and a revenue sharing and operational collaboration module. Data acquisition and fusion module: used for multi-source heterogeneous data access, cleaning, and fusion; within each time window Within this process, raw data is acquired, including real-time data from external business sources for K flights and M flights; and data initialization is completed within a preset time to obtain a confidence feature vector. ; Load forecasting module: used for macro-level passenger flow forecasting based on machine learning; defines time windows. , confidence feature vector Input into the load forecasting model and obtain the time window Total demand Total available transport capacity and load rate ; Customer segmentation and intention prediction module: used for micro-level modeling of individual customer purchase intentions; obtaining time windows. Any candidate passenger eigenvectors ; will passengers Features Input into the wide-depth hybrid prediction model, the model outputs passenger data. Probability of immediate purchase given a price. ; Obtain passenger Price sensitivity and rigid demand, and will passengers Classify customers into different types and assign them a discrete label. The subsequent pricing is guided by the segmentation logic; Dynamic pricing and demand processing module: A decision engine that combines load rate and willingness probability; based on real-time system load rate. With passengers Purchase probability distribution Automatic solution time window Best fare or discount level ; Revenue Sharing and Operational Coordination Module: Used for profit sharing and dynamic capacity scheduling; Passengers After a transaction is successful, a two-step pruning approximate Shapley value algorithm is used to calculate and attribute the benefits to each party in real time. When the future time window The system load prediction meets the continuous load condition. At that time, the system generates a capacity suggestion vector and sends an early warning to the operations dispatch center.

8. The system for dynamic processing of intermodal ticketing based on load forecasting and implicit customers according to claim 7, characterized in that, The data acquisition and fusion module includes a raw data credibility scoring module, a raw data quality identification module, and a raw data repair module; The system assigns a raw data credibility score to each data source to determine its reliability. Calculate credibility score Defined as: （1） in, The timeliness of the data is defined as follows: , Indicated as source The latest data update time is far from the time window Time difference, The time decay constant; Represented as a time window The first Data sources; This indicates the consistency between each data source and the system reference baseline; it is defined as: (2) in, It is the historical average, the weighted average of the same indicator from other sources, or the smoothed value of the previous time window; To prevent tiny constants with a denominator of 0; , For the weighting coefficients, satisfying + =1; The raw data quality assessment module uses a global threshold. Examine the data quality from each source, if If so, the data from that source is considered reliable; like If the data is unreliable, it should be corrected. For low-reliability data, the system will... Perform a fast correction; the correction method is a weighted average with historical smoothed values. （3） in, It is a smoothed reference value from the previous time window or the same time period of the previous day; The value is set based on business importance and scenario reliability; The raw data repair module is used to stitch together reliable or corrected data from all sources to obtain a time window. Confidence feature vector All sources refer to the system within the time window. The data sub-streams obtained from external modules include flight data, bus data, ticketing data, and weather data; Data concatenation, simply put, is the process of linking the data vectors from different sources together to form a larger input vector. （4）。 9. The system for dynamic processing of intermodal ticketing based on load forecasting and implicit customers according to claim 7, characterized in that, Obtain the prediction time window using the load forecasting model Total demand Total available transport capacity and load rate The specific details are as follows: Define time window The time interval covered is: , Indicates the duration of the window, from which the window begins to open. Beginning, to Finish; Define time window Available transport capacity for: （5） in, For time window Number of departures This refers to the number of seats per vehicle. The duration of the time window; Time window based on machine learning model Demand forecast Defined as: （6） Where A is the confidence feature vector output in step 1. Describe the time window Environment state; function It is a trained prediction model that has the ability to process multidimensional nonlinear relationships, heterogeneous features and time series data. The model can be selected from gradient boosting tree model, random forest, multilayer perceptron, deep neural network, long short-term memory network or gated recurrent unit. Time window The load rate is: （7） The customer segmentation and willingness prediction module includes a customer segmentation module and a willingness prediction module; Customer segmentation module, based on ticket-purchasing passengers eigenvectors Input into a lightweight wide-depth mixture prediction model, ticket-buying passengers eigenvectors for: （8） in, This refers to user profiles, with typical fields including age, place of residence, travel frequency, and average advance ticket purchase time. Refers to the time window Environmental changes, including holidays, weather, and flight density; Time window Internal load rate; This refers to the compatibility between a user's historical travel routes and their target route. Refers to the most recent n ticket purchase records and payment channels; The probability of the passenger making an immediate purchase under a given price condition, output by the wide-depth hybrid prediction model, is denoted as: （9） in, Indicates passenger In the time window The event of completing the ticket purchase within the specified time; It refers to a specific candidate discount or offer within the system; Then inform the model of this tourist Who is it, what is the current situation, and what price are we offering? Based on real-time purchase probability Passengers The system automatically categorizes customers into three types: existing customers, wavering customers, and low-interest customers; and provides services for travelers. In the time window Assign a discrete label ; The willingness prediction module makes predictions based on customer categories: Existing customers still have a high probability of making a purchase even without discounts: , Taking experience as an example, it is believed that passengers They will generally buy unconditionally, so discounts have little impact on them; their existing customers are those with essential needs. Uncertain customers are less likely to buy when there is no discount, but their purchase probability increases significantly when there is a moderate discount: , in Discounts are the threshold for driving volume growth; swing customers are price-sensitive customers; Low-intention customers: These are customers who, even with significant discounts, are still less likely to purchase than a certain threshold. Low-willingness customers are weakly related customers; in, , , Thresholds for classifying customer preferences; The threshold for increasing purchase volume driven by discounts.

10. The system for dynamic processing of intermodal ticketing based on load forecasting and implicit customers according to claim 7, characterized in that, The dynamic pricing and demand processing module also includes an environment status input unit, an action selection unit, an execution and feedback unit, a reward and revenue unit, and a strategy update unit; The environment status input unit is used to input status information including system load rate, customer profile vector, and time information; system load rate Based on the predicted demand With total available capacity Calculations show that The target passenger group profile aggregation vector is generated by statistically analyzing the characteristics of all passengers using an aggregation function. The system also utilizes the load rate to construct a discount cap constraint function: （10） in, Indicates the maximum allowed discount; For the Sigmoid function; Indicates adjustment of steepness; Indicates the center point of the Sigmoid function; The action selection unit is the agent's choice of a discount action from a predefined discount set {0%, 10%, 20%, 30%}. ; The selected action must meet the following requirements: The discount must be within the allowed range for candidate discounts. The corresponding bus fare is calculated as follows: ,in, The base fare; The execution and feedback unit is where the system quotes a price based on the selected discount, obtains the user's purchase result and actual revenue; the probability of a single passenger purchasing is... During the time window The average purchase probability for domestic travelers is: （11） The expected number of actual transactions is: （12） Discount The expected return is: （13） in, This is a cost function related to discounts. Discount-related costs include discount concessions and channel promotion expenses, defined as: , This is the discount cost coefficient; the actual and expected returns will be entered into the reward unit. The reward unit calculates the reward value based on the difference between the actual reward and the maximum reward: ,in, This represents the baseline expected return with or without a discount; reward. Used to measure actions The advantages and disadvantages; The policy update unit is where the agent updates its policy parameters using Q-learning or DQN algorithms to maximize long-term returns. In the reinforcement learning DQN framework, the system finds the optimal discount level by searching or parsing an approximation within an allowed interval. （14） Simultaneously satisfying the constraints in, The minimum allowed ticket price, The maximum magnitude of the two price adjustments, For the previous time window The discount, or allowed range, refers to a feasible discount domain jointly defined by the constraints. The revenue sharing and operational coordination module includes a revenue sharing unit and an operational coordination unit; Revenue sharing unit, used for real-time calculation and attribution to all parties; settlement period is a time window. The intermodal transport transactions actually completed within the region constitute a distributable revenue pool, and the distributable net revenue of the revenue pool is... Let the set of partners who participated in this intermodal transport service and contributed to the transaction be: The value function of the participants is defined as follows: ,in, It is a subset of any set of participants. This indicates if only this group of participants is collaborating; the classic Shapley value represents each participant. Define the weighted average of its marginal contributions: （15） When the Fang joins any existing group of partners The expected value of the gains brought by the addition of participants in the system is calculated over all possible joining orders. The order in which participants join the system is random, and the contribution of the same participant will change with different joining orders. Therefore, it is necessary to calculate the expected value over all joining orders, i.e., to calculate the average value. A two-step pruning approximation Shapley value algorithm is used as both the contribution pruning algorithm and the alliance value pruning algorithm. The algorithm content is as follows: The system employs contribution-based pruning, first estimating the upper bound of the marginal contribution of each participant. : （16） in, This represents the potential revenue generated when all stakeholders in the system participate. Indicates the removal of the main body The benefits that the system can generate; Then the participating parties are considered to be within the time window. The marginal effect of those involved is minimal and they are not included in the real-time amortization calculation set, resulting in a reduced set of core participants: ；（17） in, The contribution threshold; In the core stakeholder group For the remaining few core participants, the system uses Monte Carlo sampling to estimate the Shapley value, obtaining an approximate Shapley contribution for each retainer, denoted as . Ultimately, each retainer was obtained. Weighting allocation: (18) According to the time window Distributable net income of the income pool Divide proportionally: （19） in, The participants are within the time window The corresponding settlement amount; The operational coordination unit is used by the system to plan for future revenue sharing. The load rate prediction is calculated for each consecutive time window: Set a capacity strain threshold When satisfied It is a natural number, meaning the system's predicted value remains under high load; the system generates capacity intervention suggestions; defined as a suggestion vector: ，（20） in, For the suggested increase in the number of classes, For the suggested increase in seating capacity, High-risk routes are numbered. Specific time windows for the proposed addition .

Images