Shared bicycle unmet demand prediction method and system based on lower bound deletion learning

Abstract

The application discloses a shared bicycle unmet demand prediction method and system based on lower bound deletion learning, and relates to the technical field of shared bicycle operation scheduling. The method is based on the station vehicle in station state record of the shared bicycle system, and the observation demand of each station in each time unit is counted to identify the supply limited unit that cannot meet the demand for taking a bicycle due to no available vehicle. The observation demand of the supply limited unit is taken as the lower bound of the true demand, and the observation demand of the non-limited unit is taken as the accurate observation of the true demand, so as to construct the lower bound deletion sample and the accurate observation sample. After the construction of the feature vector, the lower bound deletion regression model is established to restore the true demand, and the predicted true demand is not lower than the observation demand. Finally, the difference between the predicted true demand and the observation demand is taken as the non-negative predicted unmet demand output. The application can predict the true demand only by the in-station state record without the failure taking request log, and can avoid underestimating the true demand in the empty period.

Description

Technical Field

[0001] This invention relates to the field of shared bicycle operation and scheduling technology, specifically to a method and system for predicting unmet demand for shared bicycles based on lower bound censoring learning. Background Technology

[0002] Bike-sharing systems typically rely on an operating platform to automatically record operational data such as vehicle ID, station ID, vehicle status at the station, station latitude and longitude, vehicle type, and timestamp. This operational data differs from ordinary order logs: ordinary order logs only reflect successfully completed bike pick-ups, while vehicle status records indicate whether a particular station has available bikes at a given timestamp, thus identifying empty bike slots at a station. Based on this data, the actual number of bike pick-ups at a particular station within a specific time period can be tallied and further used for demand forecasting, vehicle dispatching, station replenishment, and service level assessment.

[0003] However, existing shared bike operation data typically only records successful bike retrievals. When a station has no available bikes, even if a user needs to ride, they cannot retrieve a bike at that station, and no corresponding successful retrieval record will be generated in the platform's regular order data. Therefore, the observed order volume during periods of station vacancy does not equate to the actual demand, but rather is a censored result due to vehicle supply constraints. Directly using this observed order volume as the basis for prediction or analysis of actual demand can easily underestimate the true demand during peak hours, at high-demand stations, and in areas with insufficient supply.

[0004] Existing methods for forecasting shared bicycle demand mostly use historical order volumes, weather, time, and spatial characteristics as inputs to predict future rental or return volumes. While some methods consider issues such as site imbalances, supply shortages, or demand correction, they typically employ empirical correction, XGBoost correction, scheduling simulation, Poisson arrival, machine learning regression, or conventional data completion methods. These methods either rely on manually set correction coefficients, treat empty warehouse hours as ordinary missing values ​​for imputation, or only perform macroscopic corrections to the total volume, failing to establish a censored learning prediction mechanism that uses "observed demand as a lower bound of actual demand" for site empty warehouse hours. Even in existing technologies that use Gaussian censored regression models (such as the Gaussian Tobit model) to model site demand, because these methods typically perform censored regression directly at the original demand scale without imposing a lower bound constraint on the predicted results, they are still prone to systematic underestimation, even outputting predicted values ​​lower than the observed order volumes for that site, especially when shared bicycle demand exhibits a non-negative count, a large number of zero values, and a right-skewed distribution.

[0005] In summary, the existing technology has the following drawbacks: (1) Ordinary order data only records successful vehicle pick-up behavior. The observed order volume during the empty warehouse hours of the station is the result of censoring after being constrained by vehicle supply. Directly using it as the real demand for prediction or analysis will systematically underestimate the real demand during peak hours, high-demand stations and areas with insufficient supply; (2) The existing demand prediction and correction methods do not establish a censoring learning prediction mechanism for the empty warehouse hours of the station, which is based on "observed demand as the lower bound of real demand". It is easy to mistake the low observed orders during the empty warehouse hours as low real demand; (3) The underestimation of real demand will further affect the vehicle replenishment decision, vehicle deployment and service capacity assessment. Moreover, under the conditions of no failed vehicle pick-up request logs, no user search logs and no complete user intent records, there is a lack of directly observable sources of unmet demand data. (4) Even if Gaussian censored regression (e.g. Gaussian Tobit model) is used to model the site demand, since it usually performs censored regression directly on the original demand scale and does not impose the lower bound constraint of the observed demand on the prediction results, it will still produce systematic underestimation when the demand presents a non-negative count, a large number of zero values ​​and a right-skewed distribution, and may even output a predicted value lower than the observed order volume of the site, making it difficult to guarantee the compatibility of the prediction results with the observed lower bound.

[0006] Therefore, it is necessary to propose a method for predicting unmet demand for shared bicycles during empty periods at bike-sharing stations. This method can use the observed demand during empty hours as a lower bound of the actual demand without directly observing failed bike-sharing requests. It can then predict the actual demand based on a lower bound censoring learning model and output a non-negative estimate of the unmet demand. Summary of the Invention

[0007] To address the aforementioned shortcomings of existing technologies, the technical problem this invention aims to solve is: how to identify empty bike-sharing periods at shared bike stations by utilizing the station's vehicle status records under conditions of no failed bike-sharing request logs, no user search logs, and no complete user intent records; how to use the observed demand during empty bike-sharing periods as a lower bound for the actual demand for censoring learning; and how to predict the actual demand at the station and hourly scales and output non-negative estimates of unmet demand to avoid underestimating the actual demand during empty bike-sharing periods.

[0008] To address the aforementioned technical problems, the first aspect of this invention provides a method for predicting unmet demand for shared bicycles based on lower bound censoring learning. The method predicts demand based on pre-formed station vehicle status records in the shared bicycle system. Each record in the station vehicle status records corresponds to a vehicle being observed as parked at a certain station at a certain timestamp, and includes at least the vehicle number, station number, and timestamp. The method includes the following steps: Step 1: Based on the vehicle status records at the stations, calculate the observation demand for each station in each time unit. The observation demand is the number of vehicles that the station is expected to pick up in that time unit. Step 2: Based on the vehicle status records at the stations, identify whether each station is in a supply-constrained state in each time unit. The supply-constrained state is a state in which there are no available vehicles at the station and the demand for picking up vehicles cannot be met. Stations and time units in the supply-constrained state are recorded as constrained units, and the rest are recorded as unconstrained units. Step 3: Use the observation requirements of the unrestricted units as the precise observations of the potential true requirements of the unit to construct precise observation samples; use the observation requirements of the restricted units as the lower bound of the potential true requirements of the unit to construct lower bound censored samples, wherein the potential true requirements of the lower bound censored samples are not less than their observation requirements. Step 4: Construct a feature vector for each site and time unit. The feature vector shall include at least site attribute features and time attribute features. Step 5: Establish a lower bound censoring regression model based on the precise observation samples and the lower bound censored samples. The lower bound censoring regression model is used to map the feature vector to potential real demand, and the precise observation samples are modeled according to precise observation, and the lower bound censored samples are censored according to their lower bound. Step 6: For the restricted unit to be predicted, predict the potential real demand of the restricted unit based on the lower bound censoring regression model to obtain the predicted real demand, and ensure that the predicted real demand is not less than the observed demand of the restricted unit. Step 7: The difference between the predicted actual demand and the observed demand of the restricted unit is taken as the predicted unmet demand of the restricted unit, and the predicted unmet demand is a non-negative value; for unrestricted units, the predicted unmet demand is zero; output the predicted actual demand and the predicted unmet demand.

[0009] In this invention, the time unit can be an hour or other time granularities of equal length; the supply-constrained state corresponds to the empty warehouse state of a station in the shared bicycle scenario.

[0010] Furthermore, the time unit is an hour, and a station and hour demand grid is constructed using station s and hour t as indices; in step 1, valid rides are identified based on the station vehicle on-site status records, and the departure events of valid rides are aggregated using departure station s and hour t as indices. The number of valid rides departing from station s within hour t is taken as the observation demand for that station and hour. In step 2, based on the station vehicle status records, the continuous empty warehouse intervals of each station are identified. The continuous empty warehouse interval is a continuous period of time during which the station has no available vehicles. Stations and hourly units that overlap with the continuous empty warehouse intervals are marked as empty warehouse hours, and empty warehouse markers are recorded. As the restricted unit; mark the remaining stations and hourly units as non-empty hours, and record the empty hour markers. As the unrestricted unit; in step 3, the potential real demand With the aforementioned observation requirements The relationship between them is: .

[0011] Furthermore, the station vehicle status record also includes the station longitude and station latitude; in step 1, identifying a valid ride includes: sorting the station vehicle status records of the same vehicle by time according to the vehicle number; calculating the spatial distance between two adjacent records based on the station longitude and station latitude; and comparing the station position change, the spatial distance, and the time interval between two adjacent records; when two adjacent records of the same vehicle show that they are still at the same station, or the spatial distance between two adjacent records is less than a preset distance threshold, or the time interval between two adjacent records is less than a preset time threshold, the two adjacent records are considered to be in the same stationary parking state and are not identified as a valid ride; when two adjacent records of the same vehicle show that they have changed from one station to another, and the spatial distance between two adjacent records is not less than the preset distance threshold and the time interval is not less than the preset time threshold, the change in vehicle status is identified as a valid ride, the station corresponding to the previous record is taken as the starting station of the valid ride, and the station corresponding to the next record is taken as the arrival station of the valid ride; the preset distance threshold is 20~60 meters, and the preset time threshold is 20~90 seconds.

[0012] Further, in step 4, the feature vector includes station number, hour, weekday, month, whether it is a weekend, and historical adjacent time period demand statistics. The station number is the station attribute feature, and the hour, weekday, month, and whether it is a weekend are the time attribute features. The historical adjacent time period demand statistics include the mean, standard deviation, and maximum demand of station s within a preset time window around hour t. The preset time window is an hourly time period centered on hour t and within a preset number of historical days that is the same as hour t. When constructing the historical adjacent time period demand statistics, the observed demand corresponding to empty warehouse hours is excluded. .

[0013] Furthermore, in step 5, before establishing the lower bound censored regression model, the observation requirements are... and the aforementioned potential real needs Perform a log(1+y) transformation to obtain the observation requirements after the transformation. and potential real demand at the log scale , where log(·) represents the natural logarithm function; the lower bound censored regression model is established on the log(1+y) scale.

[0014] Furthermore, the lower bound censored regression model will represent the potential real demand on the log scale. This can be represented as the latent Gaussian demand process determined by the eigenvectors: in, Representing the eigenvector transpose, This represents the vector of regression parameters to be estimated. Represents the random error term. This indicates that the mean is 0 and the variance is 0. The normal distribution This represents the standard deviation of the error term; for non-empty hours, For hours when there is no trading, Estimate the regression parameter vector based on the precise observation samples and the lower bound censored samples. and the standard deviation of the error term In step 6, for the empty trading hours to be predicted, the unconstrained predicted mean of the empty trading hours on the log scale is calculated according to the lower bound censored regression model. : Calculate the standardized censoring threshold : In satisfying Under the given conditions, the conditional expected value of the potential real demand for that empty warehouse hour on the log scale is calculated using the truncated normal conditional expected value. : in, This represents the probability density function of the standard normal distribution. This represents the cumulative distribution function of the standard normal distribution. Represents the inverse Mills ratio, used for the unconstrained predicted mean under lower bound censoring conditions. Make an upward correction.

[0015] Furthermore, in step 6, the prediction of actual demand... The calculation formula is: Where exp(·) represents the natural exponential function, round(·) represents the rounding function, and max{·} represents the maximum value function.

[0016] Furthermore, the regression parameter vector and the standard deviation of the error term The log-likelihood function of the maximum likelihood estimation is obtained through maximum likelihood estimation: The first summation term represents the log-likelihood of the exact observations in the non-empty warehouse, while the second summation term represents the log-likelihood of the censored samples in the empty warehouse. This represents the probability density function of the standard normal distribution. The cumulative distribution function represents the standard normal distribution; or, the regression parameter vector β and the standard deviation σ of the error term are estimated iteratively using the expectation-maximization algorithm. In the expectation step of each iteration, the truncated normal condition expectation and truncated normal condition variance of the potential real demand on a logarithmic scale are calculated for the censored samples at the lower bound of empty positions. In the maximization step, the truncated normal condition expectation is used as the imputation value for the sample, and the regression parameter vector β is updated together with the precisely observed samples at the non-empty positions. When updating the standard deviation σ of the error term, in addition to the sum of squares of the residuals of each sample, the sum of the truncated normal condition variances of all censored samples at the lower bound of empty positions is also included. The expectation step and the maximization step are executed alternately until convergence.

[0017] Furthermore, the method also includes: aggregating the predicted unmet demand according to hour, date, station, administrative region or operating area to form at least one of the following: empty warehouse hotspot table, hourly unmet demand curve, station replenishment priority list and regional demand underestimation identification result.

[0018] In this invention, the site attribute features include a site number, and the time attribute features include hour, day of the week, month, and whether it is a weekend; the site attribute features and the time attribute features may also include other site attribute information and time attribute information that can reflect the spatiotemporal context of site and hourly demand. The feature vector This invention describes the spatiotemporal context of site and hourly demand, enabling the lower bound censoring learning model to estimate potential real demand based on site attributes, time attributes, and historical demand patterns. The core of this invention does not lie in employing a specific machine learning model, but rather in using empty warehouse hours as lower bound censoring samples and performing demand forecasting under lower bound constraints.

[0019] To address the aforementioned technical problems, a second aspect of this invention provides a shared bicycle unmet demand prediction system based on lower bound censoring learning. The system predicts unmet demand based on pre-formed station vehicle status records of the shared bicycle system. Each record in the station vehicle status records corresponds to a vehicle being observed as parked at a station at a specific timestamp, and includes at least the vehicle number, station number, and timestamp. The system includes an observation demand statistics module, a supply constraint identification module, a sample construction module, a feature construction module, a censoring modeling module, a demand restoration module, and an unmet demand output module, wherein: The observation demand statistics module is used to count the observation demand of each station in each time unit based on the station vehicle status record. The observation demand is the amount of vehicle pick-up demand observed at the station in that time unit. The supply-constrained identification module is used to identify whether each station is in a supply-constrained state in each time unit based on the station vehicle status record. The supply-constrained state is a state in which the station has no available vehicles and cannot meet the vehicle pick-up demand. The station and time unit in the supply-constrained state are recorded as constrained units, and the other stations and time units are recorded as unconstrained units. The sample construction module is used to construct accurate observation samples by taking the observation requirements of unrestricted units as accurate observations of the potential true requirements of the units; and to construct lower bound censored samples by taking the observation requirements of restricted units as lower bounds of the potential true requirements of the units, wherein the potential true requirements of the lower bound censored samples are not less than their observation requirements. The feature construction module is used to construct a feature vector for each site and time unit, and the feature vector includes at least site attribute features and time attribute features; The censoring modeling module is used to establish a lower bound censoring regression model based on the precise observation samples and the lower bound censored samples. The lower bound censoring regression model is used to map the feature vector to potential real demand, and to perform censoring modeling on the precise observation samples according to precise observation and on the lower bound censored samples according to their lower bound. The demand restoration module is used to predict the potential real demand of a restricted unit based on the lower bound censoring regression model, obtain the predicted real demand, and ensure that the predicted real demand is not less than the observed demand of the restricted unit. The unmet demand output module is used to take the difference between the predicted actual demand and the observed demand of the restricted unit as the predicted unmet demand of the restricted unit, wherein the predicted unmet demand is a non-negative value and the predicted unmet demand of the unrestricted unit is zero; and outputs the predicted actual demand and the predicted unmet demand.

[0020] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention explicitly defines the observed demand during empty hours as the lower bound of potential real demand, and uses the observed demand during non-empty hours as the precise observation of potential real demand. It incorporates both types of samples into the lower bound censoring regression model, instead of treating empty hours as ordinary missing values ​​or ordinary machine learning prediction samples. This avoids mistaking low observed orders during empty hours for low real demand and avoids underestimating the real demand during peak hours, high-demand sites, and areas with insufficient supply. (2) This invention does not require failed vehicle retrieval request logs, user search logs or complete user intent records. It can complete the prediction of real demand and unmet demand at the station and hour scale based solely on the station vehicle status records commonly available in shared bicycle systems. It has a wide range of applications and strong data availability. (3) The present invention uses log(1+y) transformation, which can retain the valid definition when the observation demand is 0 and reduce the right skewness of the demand count data, so that the non-empty warehouse accurate observation samples and the empty warehouse lower bound censored samples can be processed in a unified model. (4) The present invention adopts the truncated normal conditional expectation and corrects the unconstrained prediction mean upward by the inverse Mills ratio. It can utilize the constraint that "the actual demand is not lower than the observed demand" to avoid directly taking the low observed order volume of empty warehouse hours as the actual demand. (5) This invention uses inverse transformation and maximum value constraint post-processing to ensure that the predicted actual demand is always not less than the observed demand, and the predicted unmet demand is always non-negative, thus avoiding the unreasonable result of negative unmet demand or downward correction of the observed order quantity. (6) The output results of this invention can be aggregated by hour, date, station, administrative region or operating area to form an empty warehouse hot spot table, hourly unmet demand curve, station replenishment priority list and regional demand underestimation identification results, which can be directly used for shared bicycle vehicle scheduling, station capacity assessment, replenishment resource allocation and service level monitoring. (7) This invention overcomes the obstacle of applying the closed-form Gaussian reduction method to count-type lower bound censoring requirements: The log(1+y) transformation serves as a bridging feature, simultaneously satisfying the conflicting constraints of three terms: stable variance of the count requirement, defined and zeroed at the observation requirement of 0, and strict monotonicity. This allows the truncated normal conditional expectation reduction, which only holds under continuous homoscedastic Gaussian latent variables, to be applied to non-Gaussian, heteroscedastic, and count requirements containing a large number of zero values, while maintaining the lower bound relationship on a logarithmic scale. It overcomes the problems of undefined values ​​at zero requirements and the destruction of censoring direction in non-monotonic transformations; (8) After completing the truncated normal condition expectation restoration on the logarithmic scale, the present invention uses maximum value constraint post-processing to take the maximum value of the inverse transformation result and the observation requirement, ensuring that the predicted real requirement is given in integer form and is not lower than the observation requirement; when using truncated normal condition expectation restoration, this post-processing further ensures the lower bound compatibility of the inverse transformation and the rounding stage. (9) The log(1+y) transformation and the truncated normal conditional expectation restoration of the present invention work together synergistically: without the log(1+y) transformation, the closed Gaussian truncated normal restoration is difficult to adapt to non-negative integer, zero-inflation, and right-skewed count-type demand data; without the truncated normal conditional expectation restoration, it is impossible to restore the true demand from the lower bound censored samples of empty hours. It is the combination of the log(1+y) transformation and the truncated normal conditional expectation restoration that overcomes the combination obstacles in mechanism, parameters and working conditions between the closed Gaussian truncated normal restoration (for continuous, homoscedastic, Gaussian latent variables) and the count-type lower bound censored demand of shared bicycle stations and hours (non-negative integer, zero-inflation, heteroscedastic, censoring threshold endogenous) in terms of mechanism, parameters and working conditions, so that the closed, interpretable Gaussian truncated normal restoration can be applied to this type of count-type lower bound censored demand, and can complete the prediction of true demand and unmet demand under the condition of no failed car pick-up request log; this combination is the key difference between the present invention and the existing closed Gaussian censored restoration method. Based on this, the maximum value constraint post-processing ensures that the predicted actual demand is not lower than the lower bound of the observed demand in the form of explicit hard constraints, so as to maintain the lower bound compatibility and non-negativity of unmet demand even when numerical rounding, model approximation or other prediction value methods are used.

[0021] (10) Compared with the approach of directly performing Gaussian Tobit censoring regression on the original demand scale without imposing observation lower bound constraints, the present invention combines log(1+y) transformation with post-processing of maximum lower bound constraints. This effectively suppresses systematic underestimation when demand presents non-negative counts, a large number of zero values ​​and a right-skewed distribution, and rigidly guarantees that the predicted real demand is not lower than the observed order volume of the site. This significantly improves the compatibility between the prediction results and the observed lower bound, and further improves the restoration accuracy. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the following description of the embodiments are briefly introduced. The accompanying drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a method for predicting unmet demand for shared bicycles based on lower bound censoring learning, provided in an embodiment of the present invention. Figure 2This is a structural block diagram of a shared bicycle unmet demand prediction system based on lower bound censoring learning, provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0025] Example 1 like Figure 1 As shown, this embodiment provides a method for predicting unmet demand for shared bicycles based on lower bound censoring learning. The method uses pre-generated station vehicle availability records from the shared bicycle system as input. These records are vehicle parking status data recorded by the operating system within a preset sampling interval. Each record corresponds to a specific vehicle being observed parked at a specific station at a specific timestamp, and includes at least the vehicle number, station number, station longitude, station latitude, vehicle type, and timestamp. In the case of electric-assisted bicycles, a vehicle battery level field may also be included. The input data reflects whether available vehicles exist at each station at different timestamps and does not require the inclusion of complete riding trajectories, user search logs, or failed bike retrieval request logs.

[0026] It should be noted that the vehicle type and vehicle battery level are optional background fields. The vehicle type and vehicle battery level are not used in the steps, formulas, feature vectors, and demand restoration process of this embodiment. Their presence or absence does not affect the statistical analysis of observed demand, the identification of supply-constrained units, and the restoration of real demand in this solution. When necessary, the vehicle type can be introduced as an additional stratification basis or feature, and the vehicle battery level can be introduced as a further condition for determining vehicle availability, but neither of these is a necessary component of this solution.

[0027] The method includes the following steps 1 to 7.

[0028] Step 1: Calculate the observation requirements of each station within each hour: Based on the pre-formed station vehicle status records, sort consecutive records of the same vehicle by vehicle number and compare the station location changes, spatial distance, and time interval between adjacent records. When two adjacent records of the same vehicle show that they are still at the same station, or the spatial distance between two adjacent records is less than a preset distance threshold, or the time interval between two adjacent records is less than a preset time threshold, the two adjacent records are considered to be in the same stationary parking state and are not recognized as valid rides. When two adjacent records of the same vehicle show that they have changed from one station to another, and the spatial distance between the two adjacent records is not less than the preset distance threshold and the time interval is not less than the preset time threshold, the change in vehicle status is recognized as a valid ride. The station corresponding to the previous record is taken as the starting station of the valid ride, and the station corresponding to the next record is taken as the arrival station of the valid ride.

[0029] In this embodiment, the preset distance threshold is 30 meters, and the preset time threshold is 30 seconds. These thresholds are used to eliminate duplicate records, positioning errors, and abnormal short-term state changes, avoiding misjudging minor disturbances of the vehicle within the station as valid rides.

[0030] It should be noted that the time interval between two adjacent on-site records of the same vehicle is not less than the preset sampling interval. Therefore, the preset time threshold should not be less than the preset sampling interval; otherwise, the time threshold criterion will not be triggered. When the sampling interval of the operating system is long, the preset time threshold should be increased accordingly (see Example 2). The preset sampling interval is usually 1 second to 60 seconds.

[0031] It should be noted that in the most basic implementation, valid rides can be identified solely based on changes in station numbers: when two adjacent records of the same vehicle show a change in their station numbers, this change is identified as a valid ride. In this case, the station vehicle status record does not need to include station longitude and latitude; only the vehicle number, station number, and timestamp are required to implement steps 1 to 7 of the present invention. This embodiment further introduces station longitude and latitude to calculate the spatial distance between two adjacent records, and combines this with a preset distance threshold to exclude misjudgments caused by positioning drift between adjacent stations with similar locations, thereby improving the accuracy of valid ride identification. Accordingly, the station longitude and latitude are only used for valid ride identification in step 1; steps 2 to 7, the formulas, feature vectors, and the requirement restoration process in this embodiment do not use the station longitude and latitude.

[0032] After extracting valid rides, the departure events of valid rides are aggregated using the departure station s and hour t as indexes. The number of valid rides departing from station s within hour t is denoted as follows: And this will be used as the site's hourly observation requirement. The aforementioned observation requirements This represents the number of successful vehicle pickups at that station within a given hour, but does not represent the full potential demand under empty vehicle conditions.

[0033] In this embodiment, considering that the station vehicle status records are snapshots formed according to the sampling period, the actual departure time of a valid ride lies between the last station record of the vehicle being parked at the departure station and the next record. This embodiment uses the timestamp of the previous station record of the valid ride—that is, the last time the vehicle was observed to be parked at the departure station—as the departure time of the departure event, and the hour of that departure time is used as the hour t to which the departure event belongs. In other embodiments, the departure time inferred from the timestamps of the previous and next station records, for example, the midpoint between them, can also be used as the departure time, and the hour of that midpoint can be used as the hour t to which the departure event belongs. The above time allocation rules for departure events are consistent across all stations and all valid rides, thus uniquely determining the statistical caliber of the observation requirements.

[0034] Step 2: Identify station empty hours as supply-constrained units: Based on the effective rides obtained in Step 1 and the original station vehicle availability records, continuously analyze the vehicle availability status of each station. When a vehicle leaves station s, if there are no other vehicles recorded at station s, the departure time of the vehicle is identified as the start time of the empty interval. The departure time is taken as the departure time of the effective ride determined in Step 1. Subsequently, find the time when the next vehicle arrives at station s, and identify the arrival time as the end time of the empty interval. If no next vehicle arrives at station s within the observation period, the end time of the observation period is taken as the end time of the empty interval. This yields the continuous empty intervals of station s, denoted as... , ),in This indicates the departure time of the last vehicle that caused station s to have no available vehicles. This indicates the time it takes for the next vehicle to arrive at station s and make that station available for vehicles again.

[0035] It should be noted that this embodiment uses the existence of arbitrary time overlap as the marking condition for empty hours, which is a conservative marking strategy: as long as there is no available car within that hour, the observed demand for that hour may be censored by supply constraints, and therefore it is treated as a lower bound censored sample. Since the lower bound censoring regression model adaptively determines the upward correction magnitude based on the standardized censoring threshold α(s,t), for hours with short empty durations, the observed demand is close to the potential real demand, α(s,t) is larger, the inverse Mills ratio is smaller, and the upward correction magnitude automatically converges. Therefore, this conservative marking strategy will not lead to a systematic overestimation of mildly empty hours. In other embodiments, a minimum overlap duration threshold (e.g., 5 minutes) can also be set, and only stations and hour units with an overlap duration not less than this threshold can be marked as empty hours.

[0036] It should be noted that the continuous empty warehouse interval only indicates that the station is in a supply-constrained state where there are no cars available for rent, and is used to locate the period of potential demand loss; this interval itself does not indicate that failed car pick-up requests are directly observed.

[0037] Construct a grid of site and hourly demand using site s and hour t as indices. For any consecutive empty warehouse interval... As long as it is within the hour interval If time overlap exists, the corresponding station and hour unit are marked as an empty hour; otherwise, the station and hour unit are marked as a non-empty hour. This empty hour marking can be written as: in, This indicates the empty warehouse marker for site s at hour t; This indicates that the site overlaps with the hourly unit and the continuous empty trading range, and belongs to an empty trading hour; This indicates that there is no overlap between the site and the hourly unit, and it belongs to a non-empty hour; Indicates a period of continuous short positions; This represents the time interval corresponding to hour t. Through this step, the empty warehouse status on a continuous time scale can be converted into site and hour samples that can be used for lower bound censoring modeling.

[0038] Step 3: Construct the precise observation sample and the lower bound censored sample: For During non-empty hours, the station has available vehicles within that hour, and the observed successful vehicle retrieval demand can serve as an accurate observation of actual demand; for During empty hours, when no vehicles are available for rent at the site, the observed demand cannot be considered as real demand, but only as a lower bound of real demand. Therefore, Non-empty warehouse hour observation requirements As a potential real demand Precise observations, constructing precise observation samples; The observation needs of empty warehouse hours As a potential real demand The lower bound is used to construct lower bound censored samples. The potential real needs are... With the aforementioned observation requirements The relationship between them is: in, This indicates the potential real demand for site s in hour t; This indicates the observation requirements of station s at hour t; Indicates the hour when the warehouse is not empty; This represents the empty warehouse hours. This relationship is the key limitation that distinguishes this invention from ordinary shared bicycle demand forecasting, ordinary missing value completion, and ordinary machine learning regression: this invention does not simply fill in missing values, but uses the lower bound censoring property of "actual demand is not lower than observed demand" during empty warehouse hours for prediction.

[0039] Step 4: Constructing Site and Hour Feature Vectors: Constructing feature vectors for each site and hour unit. In this embodiment, the feature vector This includes station number, hour, day of the week, month, whether it is a weekend, and historical adjacent time period demand statistics. The station number is a station attribute feature, while the hour, day of the week, month, and whether it is a weekend are time attribute features. The historical adjacent time period demand statistics include the mean, standard deviation, and maximum demand of station s within a preset time window around hour t. The preset time window is a period of hourly time centered on hour t and within a preset number of historical days, coinciding with hour t. For example, it could be a period of hourly time coinciding with hour t within 14 days before and after hour t. This method reconstructs the true demand lost due to supply constraints during empty periods from the acquired historical station vehicle attendance records. A symmetrical time window centered on hour t is used when constructing the historical adjacent time period demand statistics. To avoid contaminating features with information lost due to empty periods, the observed demand corresponding to empty hours is excluded when constructing the historical adjacent time period demand statistics. .

[0040] In this embodiment, to enable the feature vector to be input into a linear lower bound censored regression model, the various features in the feature vector are numerically encoded and concatenated as follows: For station numbers, one-hot encoding is used, mapping each station to an indicator component with a value of 0 or 1, thereby forming a fixed effect for each station in the lower bound censored regression model; for periodic features such as hours, days of the week, and months, one-hot encoding or sine and cosine encoding is used, that is, the value v of the periodic feature is mapped to a sine component sin(2πv / P) and a cosine component cos(2πv / P) according to its period length P, so as to express the period position while maintaining the periodic continuity, wherein the period length P of the hour is 24, the period length P of the day of the week is 7, and the period length P of the month is 12; for binary features such as whether it is a weekend, 0 or 1 is directly taken as the corresponding component; for continuous numerical features such as historical adjacent time period demand statistics, they are directly used as the corresponding component after standardization. The aforementioned encoded components are concatenated in a pre-agreed fixed order to form the numerical feature vector. The dimension of the regression parameter vector is the same as the dimension of the concatenated feature vector. Those skilled in the art can also use other classification feature encoding methods such as sequence number encoding or target encoding, as long as the station number and classifications such as hour, day of the week, and month can be mapped to periodic features as numerical components for use by the linear lower bound censored regression model.

[0041] The feature vector This provides a spatiotemporal context for describing site and hourly demand, enabling subsequent lower-bound censoring learning models to estimate potential real demand based on site attributes, time attributes, and historical demand patterns. It should be noted that the core of this invention does not lie in employing a specific machine learning model, but rather in using empty warehouse hours as lower-bound censoring samples and performing demand forecasting under lower-bound constraints.

[0042] Step 5: Establish a lower bound censored regression model: Since the station and hourly observation requirements are non-negative count data, and typically have a large number of zero values ​​and a right-skewed distribution, this embodiment performs a log(1+y) transformation on the requirements before modeling. For the observation requirements, the transformation is as follows: in, This indicates the observation requirements of station s after the change in hour t; Let represent the observation requirement of station s at hour t; log(·) represents the natural logarithm function. The log(1+y) transformation preserves... This provides an effective definition of time and reduces the right skewness of demand count data.

[0043] Correspondingly, the potential real demand is represented on the same transformation scale as: in, This represents the value of the potential real demand for site s on the log scale at hour t; This represents the potential actual demand at site s at hour t. By using the same transformation scale, both the exact observation samples from non-empty warehouses and the censored samples from empty warehouses can be processed in a unified model.

[0044] On the log(1+y) transformation scale, this embodiment represents the potential real demand as a potential Gaussian demand process determined by site and hourly characteristics: in, This represents the value of the potential real demand for site s on the log scale at hour t; Representing the eigenvector transpose; This represents the vector of regression parameters to be estimated; Represents the random error term; This indicates that the mean is 0 and the variance is 0. The normal distribution; This represents the standard deviation of the error term.

[0045] In this model, non-empty hours correspond to exact observations, and empty hours correspond to lower bound censoring. in, This represents the observation requirement on the log scale; This represents the value of potential real demand on the log scale. Indicates the hour when the warehouse is not empty; This represents the number of hours with no open positions. For samples with no open positions, the model uses precise observation information to estimate the parameters; for samples with no open positions, the model uses censored information with a lower bound of "potential real demand is not less than observed demand" to estimate the parameters.

[0046] In this embodiment, model parameters and The log-likelihood function obtained through maximum likelihood estimation is: By analyzing the log-likelihood function with respect to and Numerical optimization yields the maximum likelihood estimates of the model parameters. Model parameters and It can also be obtained through other parameter estimation methods applicable to censored regression models.

[0047] Step 6: Predict the actual demand of the restricted unit and ensure it is not lower than the observed demand: For the empty warehouse hours to be predicted, first calculate its unconstrained predicted mean on the log scale according to the lower bound censored regression model: in, This represents the unconstrained predicted mean of the potential actual demand for site s at hour t on the log scale. Representing the eigenvector transpose; This represents the regression parameter vector estimated by the lower bound censored regression model.

[0048] Due to the observed value of empty warehouse hours Just potential real demand To determine the lower bound, we first need to standardize the position of this lower bound relative to the unconstrained prediction mean: in, This represents the standardized censoring threshold for site s in hour t; This represents the lower bound of the observation requirement on the log scale; This represents the unconstrained prediction mean; This represents the standard deviation of the error term. It is used to measure the number of standard deviations that the lower bound of an observation deviates from the mean of an unconstrained prediction.

[0049] In satisfying Under the given conditions, this embodiment uses a truncated normal condition expectation to calculate the potential actual demand forecast for empty warehouse hours: in, This represents the conditional expected value of the potential actual demand for site s at hour t on the log scale. This represents the unconstrained predicted mean. This represents the standard deviation of the error term; This represents the probability density function of the standard normal distribution. This represents the cumulative distribution function of the standard normal distribution. Indicates the standardized censoring threshold; Represents the inverse Mills ratio, used for the unconstrained predicted mean under lower bound censoring conditions. An upward adjustment is made. This step utilizes the constraint that "actual demand is not lower than observed demand" to avoid directly treating the low observed order volume during empty hours as actual demand.

[0050] After obtaining the predicted potential actual demand at the log scale, it is inversely transformed back to the original demand scale. To avoid the predicted actual demand being lower than the observed demand, this embodiment further employs a maximum value constraint for post-processing: in, This represents the predicted actual demand for site s at hour t; This represents the observation requirement of station s at hour t; exp(·) represents the natural exponential function; This represents the conditional expected value of potential real demand on a logarithmic scale; `round(·)` denotes the rounding function; `max{·}` denotes the maximum value function. Through this process, real demand is predicted. Always not less than the observation requirements This avoids negative unmet demand or unreasonable results that would lead to downward revisions of observed order volumes.

[0051] Step 7: Calculate and output unmet needs: After completing the actual demand forecast, this embodiment uses the difference between the predicted actual demand and the observed demand as the predicted unmet needs at the station and hourly scales. in, This indicates that the forecast for site s at hour t did not meet the demand; This represents the predicted actual demand for site s at hour t; This represents the observation requirements of station s at hour t. Since step 6 has already defined... ,therefore For non-empty hours, it can be made For hours when there is no trading, This represents the estimated value of unmet demand inferred from a lower bound censoring learning model under the condition of no failed vehicle retrieval request logs.

[0052] The output of this embodiment includes at least the station number, hourly time index, and observation requirements. Empty position marker Predicting real demand and unmet demand forecasts Furthermore, the unmet demand predicted can be analyzed by hour, date, station, administrative region, or operating area. The data is aggregated to generate at least one of the following: an empty warehouse hotspot table, an hourly unmet demand curve, a site replenishment priority list, and an identification result of undervalued regional demand. These aggregated results can be used for shared bicycle vehicle scheduling, site capacity assessment, replenishment resource allocation, and service level monitoring.

[0053] Example 2 This embodiment provides another method for predicting unmet demand for shared bicycles based on lower bound censoring learning. Its overall process is the same as that of Embodiment 1, with the only differences being the following two points.

[0054] First, in step 1, the preset distance threshold is 50 meters and the preset time threshold is 60 seconds, in order to adapt to operational scenarios with large positioning errors or long sampling intervals.

[0055] Second, in step 5, the model parameters The following is obtained through iterative estimation using the EM algorithm: In step E, for each censored sample at the lower bound of the empty warehouse, the conditional expectation of its potential value on the logarithmic scale is calculated under the current parameters according to the truncated normal conditional expectation described in step 6, and its truncated normal conditional variance is calculated simultaneously. in, The inverse Mills ratio is expressed as: In step M, the conditional expectation is used as the potential value of the sample, and the regression parameter vector is updated together with the non-empty warehouse precise observation samples using ordinary least squares regression. Update the standard deviation of the error term. In addition to the sum of squared residuals for all samples, the sum of the truncated normal conditional variances for all empty warehouse lower bound censored samples must also be included, i.e.: Where N is the total number of samples. It should be noted that if only the sum of squared residuals after imputation is used for estimation in step M... Systemic underestimation This results in an upward correction to the truncated normal condition expectation. The parameters are too small, meaning the potential real demand during periods of no trading is still underestimated. Alternate between the E-step and M-step until the parameters converge.

[0056] The remaining steps, symbol definitions, and output formats of this embodiment are the same as those of Embodiment 1, and will not be repeated here.

[0057] Example 3 like Figure 2 As shown, this embodiment provides a shared bicycle unmet demand prediction system based on lower bound censoring learning. The system predicts based on pre-formed station vehicle status records of the shared bicycle system. The system includes an observation demand statistics module, a supply constraint identification module, a sample construction module, a feature construction module, a censoring modeling module, a demand restoration module, and an unmet demand output module, wherein: The observation demand statistics module is used to count the observation demand of each station in each time unit based on the station vehicle status record. The observation demand is the amount of vehicle pick-up demand observed at the station in that time unit. The supply-constrained identification module is used to identify whether each station is in a supply-constrained state in each time unit based on the station vehicle status record. The supply-constrained state is a state in which the station has no available vehicles and cannot meet the vehicle pick-up demand. The station and time unit in the supply-constrained state are recorded as constrained units, and the other stations and time units are recorded as unconstrained units. The sample construction module is used to construct accurate observation samples by taking the observation requirements of unrestricted units as accurate observations of the potential true requirements of the units; and to construct lower bound censored samples by taking the observation requirements of restricted units as lower bounds of the potential true requirements of the units, wherein the potential true requirements of the lower bound censored samples are not less than their observation requirements. The feature construction module is used to construct a feature vector for each site and time unit, and the feature vector includes at least site attribute features and time attribute features; The censoring modeling module is used to establish a lower bound censoring regression model based on the precise observation samples and the lower bound censored samples. The lower bound censoring regression model is used to map the feature vector to potential real demand, and to perform censoring modeling on the precise observation samples according to precise observation and on the lower bound censored samples according to their lower bound. The demand restoration module is used to predict the potential real demand of a restricted unit based on the lower bound censoring regression model, obtain the predicted real demand, and ensure that the predicted real demand is not less than the observed demand of the restricted unit. The unmet demand output module is used to take the difference between the predicted actual demand and the observed demand of the restricted unit as the predicted unmet demand of the restricted unit, wherein the predicted unmet demand is a non-negative value and the predicted unmet demand of the unrestricted unit is zero; and outputs the predicted actual demand and the predicted unmet demand.

[0058] The specific processing procedures of each of the above modules correspond one-to-one with steps 1 to 7 in Example 1, including effective cycling identification and its distance and time thresholds, continuous empty bay interval identification and empty bay hour marking, feature vector construction and numerical encoding of various features, log(1+y) transformation, parameter estimation of the lower bound censored regression model, truncated normal conditional expectation and inverse Mills ratio upward correction, as well as inverse transformation and maximum value constraint, etc., which will not be elaborated here.

[0059] Example 4 To verify the effectiveness of this invention in restoring real demand, a comparative experiment was conducted on the station status records of vehicles at real operating stations. The experiment selected verification samples with known real demand from real non-empty stations and hourly units, and evaluated them by constructing a pseudo-censoring method: for each verification sample, the actual observed demand was retained as the true value of the real demand, while a lower observation lower bound was artificially set to simulate the situation where only part of the demand could be observed under empty station censoring conditions, thus evaluating the ability of each method to restore real demand under the premise of known true values. The experiment compared the following four methods on the same training set and the same verification set: Method 1 is the proposed scheme, which sequentially performs log(1+y) transformation, lower bound censoring regression, truncated normal conditional expectation restoration, and post-processing by taking the maximum value constraint between the predicted value and the observation lower bound; Method 2 is the complete Gaussian Tobit, which performs Gaussian Tobit censoring regression on the original demand scale, estimates the regression parameters with maximum likelihood, and then uses truncated normal conditional expectation E[y*|y*≥y]=μ+σ·φ(α) / [1 Φ(α)] restores the potential real demand without performing a log(1+y) transformation or imposing a maximum lower bound constraint; Method 3 is a direct Gaussian Tobit, i.e., parameter estimation is the same as Method 2, but the mean x is directly predicted without constraints. T β is used to predict actual demand without truncating the expected value under normal conditions. Method four is ordinary imputation, which uses ordinary linear regression without modeling censoring. Evaluation metrics include mean absolute error (MAE), root mean square error (RMSE), bias (the mean difference between predicted actual demand and true value), and lower bound violation rate (the proportion of samples where predicted actual demand is lower than the observed lower bound). The results are stratified and statistically analyzed according to the sample-level censoring ratio (the ratio of the amount of missing data censored in a sample to its actual demand).

[0060] Table 1 Overall results of the control experiment (validation set of four cities, sample size 30628) As shown in Table 1: (i) Regarding lower bound compatibility, the lower bound violation rate of this scheme and the complete Gaussian Tobit is both 0.0%. This is because both use truncated normal conditional expectation, and their predicted values ​​are mathematically no lower than the observation lower bound. However, the direct Gaussian Tobit uses the unconstrained predicted mean as the prediction, and the lower bound violation rate is 42.2%, while the ordinary regression imputation rate is 63.1%. (ii) Regarding restoration accuracy, compared with the complete Gaussian Tobit using the correct prediction method, this scheme reduces the MAE from 1.65 to 1.61 and the RMSE from 3.24 to 3.19 by introducing the log(1+y) transformation. This further reduces the restoration error when the demand presents non-negative counts, a large number of zero values, and a right-skewed distribution. Compared with the direct Gaussian Tobit without truncated normal correction, the MAE of this scheme is reduced by about 36.7%, of which about 35 percentage points come from the standard restoration method of truncated regression using truncated normal conditional expectation and about 2 percentage points come from the log(1+y) transformation.

[0061] Table 2. Results of the controlled experiment stratified by censoring ratio (validation set of four cities) As shown in Table 2, this scheme maintains a lower bound violation rate of 0.0% and the lowest restoration error across different censoring ratios. In the censoring ratio range of 20%-50% (with an average censoring rate of approximately 33.1%), the MAE of this scheme is as low as 0.14, representing a reduction of approximately 92.7% compared to the direct Gaussian Tobit and approximately 93.5% compared to ordinary regression imputation. In the heavily censored range of greater than 50% (with an average censoring rate of approximately 65.7%), the MAE of this scheme is still approximately 32.7% lower than the direct Gaussian Tobit. This indicates that as the degree of censoring (empty positions) changes, this scheme consistently maintains a significant advantage in restoration accuracy compared to methods without modeling censoring or without applying lower bound constraints. Furthermore, the more severe the censoring, the greater the error caused by systematic underestimation in the comparative methods. This scheme effectively suppresses this systematic underestimation through lower bound censoring modeling and truncating the normal conditional expectation restoration. In the stratified statistics in Table 2, Method 2 is listed as Direct Gaussian Tobit; the lower bound violation rate of Full Gaussian Tobit is also 0.0% in each censoring ratio range, and its restoration accuracy decreases as the censoring ratio increases. The advantage of this scheme over Full Gaussian Tobit is mainly reflected in the low to medium censoring ratio range.

[0062] Further verification was conducted on vehicle station status records in four real-world operating cities (City 1, City 2, City 3, and City 4). The lower bound violation rate of this scheme was 0.0% in all four cities, and the MAE was reduced by approximately 36.1%, 27.7%, 26.9%, and 38.1% respectively compared to the direct Gaussian Tobit method. The results were consistent across all cities. The above comparative experiments show that this invention, through the combination of log(1+y) transformation, lower bound censoring regression, truncated normal conditional expectation restoration, and maximum lower bound constraint, can accurately restore the real demand in counting lower bound censoring demand scenarios and rigidly guarantee the non-negativity and lower bound compatibility of the prediction results. This represents a significant performance improvement compared to existing direct Gaussian Tobit methods and ordinary imputation methods.

[0063] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0064] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A method for predicting unmet demand for shared bicycles based on lower bound censoring learning, wherein the method predicts based on pre-formed station vehicle status records of the shared bicycle system, each record in the station vehicle status records corresponds to a vehicle being observed as parked at a certain station at a certain timestamp, and includes at least the vehicle number, station number, and timestamp, characterized in that, The method includes the following steps: Step 1: Based on the vehicle status records at the stations, calculate the observation demand for each station in each time unit. The observation demand is the number of vehicles that the station is expected to pick up in that time unit. Step 2: Based on the vehicle status records at the stations, identify whether each station is in a supply-constrained state in each time unit. The supply-constrained state is a state in which there are no available vehicles at the station and the demand for picking up vehicles cannot be met. Stations and time units in the supply-constrained state are recorded as constrained units, and the rest are recorded as unconstrained units. Step 3: Use the observation requirements of the unrestricted units as the precise observations of the potential true requirements of the unit to construct precise observation samples; use the observation requirements of the restricted units as the lower bound of the potential true requirements of the unit to construct lower bound censored samples, wherein the potential true requirements of the lower bound censored samples are not less than their observation requirements. Step 4: Construct a feature vector for each site and time unit. The feature vector shall include at least site attribute features and time attribute features. Step 5: Establish a lower bound censoring regression model based on the precise observation samples and the lower bound censored samples. The lower bound censoring regression model is used to map the feature vector to potential real demand, and the precise observation samples are modeled according to precise observation, and the lower bound censored samples are censored according to their lower bound. Step 6: For the restricted unit to be predicted, predict the potential real demand of the restricted unit based on the lower bound censoring regression model to obtain the predicted real demand, and ensure that the predicted real demand is not less than the observed demand of the restricted unit. Step 7: The difference between the predicted actual demand and the observed demand of the restricted unit is taken as the predicted unmet demand of the restricted unit. For unrestricted units, the predicted unmet demand is zero; output the predicted actual demand and the predicted unmet demand. The time unit is an hour, and a station and hour demand grid is constructed using station s and hour t as indices. In step 1, valid rides are identified based on the vehicle on-site status records of the stations. The departure events of valid rides are aggregated using departure station s and hour t as indices, and the number of valid rides departing from station s within hour t is taken as the observation demand for that station and hour. In step 2, based on the station vehicle status records, the continuous empty warehouse intervals of each station are identified. The continuous empty warehouse interval is a continuous period of time during which the station has no available vehicles. Stations and hourly units that overlap with the continuous empty warehouse intervals are marked as empty warehouse hours, and empty warehouse markers are recorded. As the restricted unit; mark the remaining stations and hourly units as non-empty hours, and record the empty hour markers. As the unrestricted unit; in step 3, the potential real demand With the aforementioned observation requirements The relationship between them is: ; In step 5, before establishing the lower bound censored regression model, the observation requirements are... and the aforementioned potential real needs Perform a log(1+y) transformation to obtain the observation requirements after the transformation. and potential real demand at the log scale , where log(·) represents the natural logarithm function; the lower bound censored regression model is established on the log(1+y) scale; The lower bound censoring regression model will represent the potential real demand on the log scale. This can be represented as the latent Gaussian demand process determined by the eigenvectors: in, Representing the eigenvector transpose, This represents the vector of regression parameters to be estimated. Represents the random error term. This indicates that the mean is 0 and the variance is 0. The normal distribution This represents the standard deviation of the error term; for non-empty hours, For hours when there is no trading, Estimate the regression parameter vector based on the precise observation samples and the lower bound censored samples. and the standard deviation of the error term For the empty trading hours to be predicted, the unconstrained predicted mean of the empty trading hours on the log scale is calculated based on the lower bound censored regression model. : Calculate the standardized censoring threshold : In satisfying Under the given conditions, the conditional expected value of the potential real demand for that empty warehouse hour on the log scale is calculated using the truncated normal conditional expected value. : in, This represents the probability density function of the standard normal distribution. This represents the cumulative distribution function of the standard normal distribution. Represents the inverse Mills ratio, used for the unconstrained predicted mean under lower bound censoring conditions. Make an upward correction.

2. The method for predicting unmet demand for shared bicycles based on lower bound censoring learning according to claim 1, characterized in that, The station vehicle status record also includes the station longitude and station latitude; in step 1, identifying a valid ride includes: sorting the station vehicle status records of the same vehicle by time according to the vehicle number; calculating the spatial distance between two adjacent records based on the station longitude and station latitude; and comparing the station position change, spatial distance, and time interval between two adjacent records; when two adjacent records of the same vehicle show that they are still located at the same station, or the spatial distance between two adjacent records is less than a preset distance threshold, or the time interval between two adjacent records is less than a preset time threshold, the two adjacent records are regarded as the same stationary parking state and are not identified as a valid ride; when two adjacent records of the same vehicle show that they have changed from one station to another, and the spatial distance between two adjacent records is not less than the preset distance threshold and the time interval is not less than the preset time threshold, the change in vehicle status is identified as a valid ride, the station corresponding to the previous record is taken as the starting station of the valid ride, and the station corresponding to the next record is taken as the destination station of the valid ride.

3. The method for predicting unmet demand for shared bicycles based on lower bound censoring learning according to claim 1, characterized in that, The feature vector includes station number, hour, weekday, month, weekend status, and historical neighboring time period demand statistics. The station number represents the station attribute feature, while the hour, weekday, month, and weekend status represent the time attribute feature. The historical neighboring time period demand statistics include the mean, standard deviation, and maximum demand for station s within a preset time window around hour t. The preset time window is an hourly time period centered on hour t and within a preset historical number of days, sharing the same hourly time period as hour t. When constructing the historical neighboring time period demand statistics, observed demand corresponding to empty warehouse hours is excluded. .

4. The method for predicting unmet demand for shared bicycles based on lower bound censoring learning according to claim 1, characterized in that, The prediction of actual demand The calculation formula is: Where exp(·) represents the natural exponential function, round(·) represents the rounding function, and max{·} represents the maximum value function.

5. The method for predicting unmet demand for shared bicycles based on lower bound censoring learning according to claim 1, characterized in that, The regression parameter vector and the standard deviation of the error term The log-likelihood function of the maximum likelihood estimation is obtained through maximum likelihood estimation: The first summation term represents the log-likelihood of the exact observations in the non-empty warehouse, while the second summation term represents the log-likelihood of the censored samples in the empty warehouse. This represents the probability density function of the standard normal distribution. The cumulative distribution function represents the standard normal distribution; or, the regression parameter vector β and the standard deviation of the error term σ are estimated iteratively using the expectation-maximization algorithm. In the expectation step of each iteration, the truncated normal condition expectation and truncated normal condition variance of the potential real demand on a logarithmic scale are calculated for the censored samples at the lower bound of empty positions. In the maximization step, the truncated normal condition expectation is used as the imputation value for the sample, and the regression parameter vector β is updated together with the precisely observed samples at the non-empty positions. When updating the standard deviation of the error term σ, in addition to the sum of squares of the residuals of each sample, the sum of the truncated normal condition variances of all censored samples at the lower bound of empty positions is also included. The expectation step and the maximization step are executed alternately until convergence.

6. The method for predicting unmet demand for shared bicycles based on lower bound censoring learning according to claim 1, characterized in that, The method further includes: aggregating the predicted unmet demand according to hour, date, station, administrative region or operating area to form at least one of the following: empty warehouse hotspot table, hourly unmet demand curve, station replenishment priority list and regional demand underestimation identification result.

7. A shared bicycle unmet demand prediction system based on lower bound censoring learning, wherein the system predicts based on pre-formed station vehicle status records of the shared bicycle system, each record in the station vehicle status records corresponds to a vehicle being observed as parked at a certain station at a certain timestamp, and includes at least the vehicle number, station number, and timestamp, characterized in that, The system includes an observation demand statistics module, a supply constraint identification module, a sample construction module, a feature construction module, a censoring modeling module, a demand restoration module, and an unmet demand output module, wherein: The observation demand statistics module is used to count the observation demand of each station in each time unit based on the station vehicle status record. The observation demand is the amount of vehicle pick-up demand observed at the station in that time unit. The supply-constrained identification module is used to identify whether each station is in a supply-constrained state in each time unit based on the station vehicle status record. The supply-constrained state is a state in which the station has no available vehicles and cannot meet the vehicle pick-up demand. The station and time unit in the supply-constrained state are recorded as constrained units, and the other stations and time units are recorded as unconstrained units. The sample construction module is used to construct accurate observation samples by taking the observation requirements of unrestricted units as accurate observations of the potential true requirements of the units; and to construct lower bound censored samples by taking the observation requirements of restricted units as lower bounds of the potential true requirements of the units, wherein the potential true requirements of the lower bound censored samples are not less than their observation requirements. The feature construction module is used to construct a feature vector for each site and time unit, and the feature vector includes at least site attribute features and time attribute features; The censoring modeling module is used to establish a lower bound censoring regression model based on the precise observation samples and the lower bound censored samples. The lower bound censoring regression model is used to map the feature vector to potential real demand, and to perform censoring modeling on the precise observation samples according to precise observation and on the lower bound censored samples according to their lower bound. The demand restoration module is used to predict the potential real demand of a restricted unit based on the lower bound censoring regression model, obtain the predicted real demand, and ensure that the predicted real demand is not less than the observed demand of the restricted unit. The unmet demand output module is used to take the difference between the predicted actual demand and the observed demand of the restricted unit as the predicted unmet demand of the restricted unit, and the predicted unmet demand of the unrestricted unit is zero; and output the predicted actual demand and the predicted unmet demand. The time unit is an hour. A grid of station and hourly demand is constructed using station s and hour t as indices. Valid rides are identified based on the vehicle on-site status records of the stations. Departure events of valid rides are aggregated using departure station s and hour t as indices. The number of valid rides departing from station s within hour t is taken as the observation demand for that station and hour. Based on the vehicle availability records at the stations, identify the continuous empty bay intervals for each station. These continuous empty bay intervals are the consecutive time periods during which no vehicles are available at the station. Stations and hourly units that overlap with these continuous empty bay intervals are marked as empty bay hours, and these empty bay markers are recorded. As the restricted unit; mark the remaining stations and hourly units as non-empty hours, and record the empty hour markers. As the unrestricted unit; the potential real demand With the aforementioned observation requirements The relationship between them is: ; Before establishing the lower bound censored regression model, the observation requirements are... and the aforementioned potential real needs Perform a log(1+y) transformation to obtain the observation requirements after the transformation. and potential real demand at the log scale , where log(·) represents the natural logarithm function; the lower bound censored regression model is established on the log(1+y) scale; The lower bound censoring regression model will represent the potential real demand on the log scale. This can be represented as the latent Gaussian demand process determined by the eigenvectors: in, Representing the eigenvector transpose, This represents the vector of regression parameters to be estimated. Represents the random error term. This indicates that the mean is 0 and the variance is 0. The normal distribution This represents the standard deviation of the error term; for non-empty hours, For hours when there is no trading, Estimate the regression parameter vector based on the precise observation samples and the lower bound censored samples. and the standard deviation of the error term For the empty trading hours to be predicted, the unconstrained predicted mean of the empty trading hours on the log scale is calculated based on the lower bound censored regression model. : Calculate the standardized censoring threshold : In satisfying Under the given conditions, the conditional expected value of the potential real demand for that empty warehouse hour on the log scale is calculated using the truncated normal conditional expected value. : in, This represents the probability density function of the standard normal distribution. This represents the cumulative distribution function of the standard normal distribution. Represents the inverse Mills ratio, used for the unconstrained predicted mean under lower bound censoring conditions. Make an upward correction.

Images