An influenza vaccine transportation optimization method and system based on data driving

Abstract

This invention provides a data-driven method and system for optimizing influenza vaccine transport, relating to the fields of data processing and intelligent optimization. The method includes acquiring multi-source dynamic data corresponding to the influenza vaccine transport task, and performing time alignment and feature construction on the multi-source dynamic data to obtain feature vectors representing a set of candidate transport links and a set of candidate transport nodes. The candidate transport links represent the transport modes and path combinations from the starting node to the destination node. This data-driven method and system for optimizing influenza vaccine transport establishes a data-driven machine learning prediction model by performing time alignment and feature construction on multi-source dynamic data such as road traffic, weather, transport node operational capabilities, and cold chain status. This model jointly predicts transport time, node waiting time, and temperature exceedance risks, achieving accurate characterization of the timeliness and cold chain risks in the influenza vaccine transport process.

Description

Technical Field

[0001] This invention relates to the field of data processing and intelligent optimization technology, specifically to a data-driven method and system for optimizing the transport of influenza vaccines. Background Technology

[0002] Existing influenza vaccine transport management typically revolves around cold chain constraints and timeliness requirements. Common practices include: route and plan development based on empirical rules or fixed parameters; vehicle routing or scheduling solutions based on deterministic optimization or heuristic algorithms; and cold chain monitoring systems based on positioning and temperature sensing for recording in-transit trajectories, recording temperature curves, and issuing alarms for exceeding limits.

[0003] However, the above-mentioned schemes generally fail to adequately depict the uncertainties in the real environment. They often treat transportation time, waiting time at key points, and temperature control risks as constants or simple rules, making it difficult to reflect dynamic factors such as traffic congestion, weather changes, loading and unloading queues, vehicle status fluctuations, and refrigerant decay, resulting in large discrepancies between plans and reality. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a data-driven method and system for optimizing influenza vaccine transport. The technical problem this invention aims to solve is: how to dynamically characterize transport time, node waiting time, and temperature control risks through an uncertainty prediction and optimization method based on multi-source dynamic data, thereby solving the problem that existing influenza vaccine transport schemes are unable to reflect real-world uncertainties and have large discrepancies between plans and actual conditions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a data-driven method for optimizing the transport of influenza vaccines, comprising: S1. Obtain multi-source dynamic data corresponding to the influenza vaccine transport task, and perform time alignment and feature construction on the multi-source dynamic data to obtain feature vectors for characterizing the candidate transport link set and the candidate transport node set. The candidate transport link is used to characterize the transport mode and path combination from the starting node to the destination node, and the candidate transport node is used to characterize the operation and waiting process of loading, unloading, handover or transfer nodes. S2. Construct a training sample set based on the feature vector, and perform model training on the training sample set to learn model parameters, thereby obtaining a machine learning prediction model. The machine learning prediction model is used to output the predicted value of transportation time, the predicted value of node waiting time, the predicted value of temperature exceeding the limit risk and its uncertainty characterization parameters. When the execution feedback data is obtained, the model parameters are updated or calibrated online based on the execution feedback data. S3. Input the influenza vaccine transport task to be planned into the machine learning prediction model, and reason about the candidate transport link set and the candidate transport node set to obtain the transport time prediction value, node waiting time prediction value, temperature exceedance risk prediction value and uncertainty characterization parameters corresponding to each candidate transport link and each candidate transport node, and summarize them to form the parameter set of the transport optimization problem. S4. Based on the parameter set, construct and solve the transfer optimization model to output the influenza vaccine transfer plan. The transfer optimization model includes cold chain constraints, transportation capacity constraints, and uncertainty constraints. The cold chain constraints include temperature threshold constraints and cold preservation capacity constraints. The transportation capacity constraints include time window constraints and capacity constraints. The uncertainty constraints constrain the uncertainty characterization parameters corresponding to the predicted transportation time, predicted node waiting time, and predicted temperature exceedance risk according to a preset confidence level, so that the influenza vaccine transfer plan satisfies the time window constraints and the cold chain constraints at the confidence level. The influenza vaccine transfer plan includes the selection results of candidate transfer links and candidate transfer nodes, the departure time and node operation period, and the vehicle and refrigerant configuration corresponding to the cold preservation capacity constraints. S5. During the execution of the influenza vaccine transport plan, execution feedback data is collected, including actual transport time, actual node waiting time and temperature data during transit. The parameters of the machine learning prediction model are calibrated or updated online based on the execution feedback data. When the prediction deviation or constraint violation risk reaches the preset trigger condition, the influenza vaccine transport plan is re-optimized.

[0006] Preferably, the multi-source dynamic data includes road traffic operation data, meteorological data, operational capacity data and queuing data of each candidate transfer node, availability and capacity data of cold chain vehicles, refrigerant inventory and replenishment data, and historical temperature data in transit. The multi-source dynamic data is uniformly mapped to a time series based on a preset time granularity before the time alignment is performed.

[0007] Preferably, the feature construction includes: constructing link-level features for the candidate transfer links and node-level features for the candidate transfer nodes; concatenating the link-level features and the node-level features in a preset feature order to generate the feature vector; the link-level features include transportation mode codes, path length, historical speed distribution parameters, meteorological impact features, and traffic congestion index; the node-level features include node operation period codes, unit time processing capacity, queue length statistics, historical waiting time distribution parameters, and node temperature control environment features.

[0008] Preferably, the training sample set is constructed from historical transfer tasks. Each training sample in the training sample set consists of a sample input and a sample output. The sample input is the feature vector extracted for the same historical transfer task within a preset observation time window. The sample output is a supervision label determined by the historical records of the historical transfer task on the candidate transfer link or candidate transfer node corresponding to the feature vector. The supervision label includes the actual transportation time, the actual node waiting time, and the temperature exceedance event marker.

[0009] Preferably, the machine learning prediction model is a probabilistic prediction model. The machine learning prediction model outputs location parameters and scale parameters as uncertainty characterization parameters for the predicted value of transportation time and the predicted value of node waiting time, and outputs risk probability value and risk confidence interval parameters as uncertainty characterization parameters for the predicted value of temperature exceeding limit risk.

[0010] Preferably, the online update or calibration includes: calculating the predicted residual based on the execution feedback data, and performing incremental learning update on the model parameters when the statistics of the predicted residual satisfy a preset drift criterion, wherein the drift criterion is a residual mean offset threshold or a residual variance increase threshold.

[0011] Preferably, the transshipment optimization model is a mixed integer optimization model that includes binary decision variables. The binary decision variables are used to represent the selection results of the candidate transshipment links and the candidate transshipment nodes. The objective function of the transshipment optimization model includes a transport time objective item and a temperature exceedance risk objective item. The objective function performs a weighted summation of each objective item using preset weights.

[0012] Preferably, the uncertainty constraint is implemented through an opportunity constraint. Under the preset confidence level, the predicted transport time and the corresponding uncertainty characterization parameter are converted into the corresponding quantile upper bound, the predicted node waiting time and the corresponding uncertainty characterization parameter are converted into the corresponding quantile upper bound, and the quantile upper bound is used to replace the deterministic duration parameter in the time window constraint, so that the influenza vaccine transport scheme satisfies the time window constraint under the confidence level.

[0013] Preferably, the preset triggering conditions include: the deviation between the actual transportation time and the predicted transportation time exceeds a first threshold; the deviation between the actual node waiting time and the predicted node waiting time exceeds a second threshold; the duration for which the temperature data in transit exceeds the temperature threshold constraint reaches a third threshold; and the predicted value of temperature exceedance risk obtained based on real-time inference exceeds a fourth threshold. When the re-optimization is triggered, the candidate transfer links and candidate transfer nodes that have been executed are locked as fixed decisions, and the transfer optimization model is re-solved for the unexecuted parts to update the departure time, node operation period, and vehicle and refrigerant configuration.

[0014] A data-driven influenza vaccine delivery optimization system includes: Feature construction module: used to acquire multi-source dynamic data corresponding to the influenza vaccine transport task, and to perform time alignment and feature construction on the multi-source dynamic data to obtain feature vectors that characterize the candidate transport link set and the candidate transport node set; Prediction model module: used to construct a training sample set based on the feature vector, and perform model training on the training sample set to learn model parameters, thereby obtaining a machine learning prediction model. The machine learning prediction model is used to output the predicted value of transportation time, the predicted value of node waiting time, the predicted value of temperature exceedance risk and its uncertainty characterization parameters, and when the execution feedback data is obtained, the model parameters are updated or calibrated online based on the execution feedback data. Parameter reasoning generation module: It is used to input the influenza vaccine transportation task to be planned into the machine learning prediction model, reason about the candidate transportation link set and the candidate transportation node set, and obtain the transportation time prediction value, node waiting time prediction value, temperature exceedance risk prediction value and uncertainty characterization parameters corresponding to each candidate transportation link and each candidate transportation node, and summarize them to form the parameter set of the transportation optimization problem. The transfer optimization solution module is used to construct and solve the transfer optimization model based on the parameter set, and output the influenza vaccine transfer plan. The transfer optimization model includes cold chain constraints, transportation capacity constraints, and uncertainty constraints. The cold chain constraints include temperature threshold constraints and cold preservation capacity constraints. The transportation capacity constraints include time window constraints and capacity constraints. The uncertainty constraints constrain the uncertainty characterization parameters corresponding to the predicted values ​​of transportation time, node waiting time, and temperature exceedance risk according to a preset confidence level, so that the influenza vaccine transfer plan satisfies the time window constraints and the cold chain constraints at the confidence level. The execution feedback collection and re-optimization module is used to collect execution feedback data during the execution of the influenza vaccine transport plan. The execution feedback data includes actual transport time, actual node waiting time and temperature data during transit. Based on the execution feedback data, the module calibrates or updates the parameters of the machine learning prediction model online, and triggers the re-optimization of the influenza vaccine transport plan when the prediction deviation or constraint violation risk reaches a preset trigger condition.

[0015] This invention provides a data-driven method and system for optimizing influenza vaccine transport. It offers the following advantages: This data-driven method and system for optimizing influenza vaccine transport establishes a data-driven machine learning prediction model by aligning and constructing features from multi-source dynamic data, including road traffic, weather, transport node operational capabilities, and cold chain status. This model jointly predicts transport time, node waiting time, and temperature exceedance risks, enabling precise characterization of the timeliness and cold chain risks in the influenza vaccine transport process.

[0016] A transshipment optimization model incorporating uncertainty constraints is adopted. Under a pre-set confidence level, a transshipment plan is generated by comprehensively considering cold chain constraints and transportation capacity constraints. Dynamic calibration and re-optimization are achieved by combining execution feedback, thereby reducing the risk of temperature exceeding limits and improving the reliability and robustness of the transshipment plan. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method of the present invention; Figure 2 This is a flowchart of the multi-source data feature construction process of the present invention; Figure 3 This is a flowchart of the machine learning prediction model training process of the present invention; Figure 4 This is a flowchart of the solution process for the transport optimization model of this invention; Figure 5 This is a flowchart of the monitoring and re-optimization process for the transfer scheme of this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 like Figure 1-5 As shown, this embodiment of the invention provides a data-driven method for optimizing influenza vaccine transport, including: S1. Acquire multi-source dynamic data corresponding to the influenza vaccine transport task, and perform time alignment and feature construction on the multi-source dynamic data to obtain feature vectors representing the candidate transport link set and candidate transport node set. The candidate transport link is used to represent the transportation mode and path combination from the starting node to the destination node, and the candidate transport node is used to represent the operation and waiting process of loading, unloading, handover or transfer nodes. The multi-source dynamic data includes road traffic operation data, meteorological data, operation capacity data and queuing data of each candidate transport node, cold chain vehicle availability and capacity data, refrigerant inventory and replenishment data, and historical in-transit temperature data. After uniformly mapping the multi-source dynamic data to a time series based on a preset time granularity, time alignment is performed. Feature construction includes: constructing link-level features for candidate transfer links and node-level features for candidate transfer nodes. The link-level features and node-level features are then concatenated in a preset feature order to generate a feature vector. The link-level features include transportation mode codes, path length, historical speed distribution parameters, meteorological impact features, and traffic congestion index. The node-level features include node operation time period codes, unit time processing capacity, queue length statistics, historical waiting time distribution parameters, and node temperature control environment features.

[0020] S2. A training sample set is constructed based on feature vectors, and the model is trained on the training sample set to learn the model parameters, thereby obtaining a machine learning prediction model. The machine learning prediction model is used to output the predicted values ​​of transportation time, node waiting time, and temperature exceedance risk, as well as their uncertainty characterization parameters. When execution feedback data is obtained, the model parameters are updated or calibrated online based on the execution feedback data. The training sample set is constructed from historical transportation tasks. Each training sample in the training sample set consists of a sample input and a sample output. The sample input is a feature vector extracted for the same historical transportation task within a preset observation time window. The sample output is a supervision label determined by the historical records of the historical transportation task on the candidate transportation links or candidate transportation nodes corresponding to the feature vectors. The supervision label includes the actual transportation time, the actual node waiting time, and the temperature exceedance event marker. The machine learning prediction model is a probabilistic prediction model. For the predicted values ​​of transportation time and node waiting time, the machine learning prediction model outputs location parameters and scale parameters as uncertainty characterization parameters, and for the predicted value of temperature exceedance risk, it outputs a risk probability value and a risk confidence interval parameter as uncertainty characterization parameters. Online updates or calibrations include: calculating the predicted residuals based on the execution feedback data, and performing incremental learning updates on the model parameters when the statistics of the predicted residuals meet the preset drift criteria. The drift criteria are the residual mean shift threshold or the residual variance increase threshold.

[0021] S3. Input the influenza vaccine transport task to be planned into the machine learning prediction model, and reason about the candidate transport link set and candidate transport node set to obtain the predicted transport time, node waiting time, temperature exceedance risk and their uncertainty characterization parameters for each candidate transport link and candidate transport node, and summarize them to form the parameter set of the transport optimization problem.

[0022] S4. Construct and solve a transport optimization model based on the parameter set to output an influenza vaccine transport plan. The transport optimization model includes cold chain constraints, transport capacity constraints, and uncertainty constraints. Cold chain constraints include temperature threshold constraints and cold preservation capacity constraints. Transport capacity constraints include time window constraints and capacity constraints. Uncertainty constraints constrain the uncertainty representation parameters corresponding to the predicted transport time, node waiting time, and temperature exceedance risk according to a preset confidence level, ensuring that the influenza vaccine transport plan meets the time window and cold chain constraints at the confidence level. The influenza vaccine transport plan includes the selection results of candidate transport links and candidate transport nodes, departure time and node operation period, and the vehicle and refrigerant configuration corresponding to the cold preservation capacity constraint. The transport optimization model is a mixed integer optimization model containing binary decision variables. The binary decision variables are used to represent the selection results of candidate transport links and candidate transport nodes. The objective function of the transport optimization model includes a transport time objective term and a temperature exceedance risk objective term. The objective function is a weighted sum of each objective term using preset weights. Uncertainty constraints are achieved through chance constraints. Under a pre-set confidence level, the predicted transport time and the corresponding uncertainty characterization parameters are converted into corresponding quantile upper bounds, and the predicted node waiting time and the corresponding uncertainty characterization parameters are converted into corresponding quantile upper bounds. The quantile upper bounds replace the deterministic duration parameters in the time window constraints, so that the influenza vaccine transport scheme meets the time window constraints under the confidence level.

[0023] S5. During the execution of the influenza vaccine transport plan, execution feedback data is collected. This data includes actual transport time, actual node waiting time, and in-transit temperature data. Based on this data, the machine learning prediction model is calibrated or updated online. When the predicted deviation or constraint violation risk reaches a preset trigger condition, the influenza vaccine transport plan is re-optimized. The preset trigger conditions include: the deviation between the actual transport time and the predicted transport time exceeds a first threshold; the deviation between the actual node waiting time and the predicted node waiting time exceeds a second threshold; the duration for which the in-transit temperature data exceeds the temperature threshold constraint reaches a third threshold; and the predicted temperature exceedance risk value obtained based on real-time inference exceeds a fourth threshold. When re-optimization is triggered, the completed candidate transport links and nodes are locked as fixed decisions, and the transport optimization model is re-solved for the unexecuted parts to update the departure time, node operation period, and vehicle and refrigerant configuration.

[0024] A data-driven influenza vaccine delivery optimization system includes: Feature construction module: used to acquire multi-source dynamic data corresponding to the influenza vaccine transportation task, and to perform time alignment and feature construction on the multi-source dynamic data to obtain feature vectors that characterize the candidate transportation link set and the candidate transportation node set.

[0025] Prediction Model Module: This module is used to construct a training sample set based on feature vectors and to train the model on the training sample set to learn the model parameters, thereby obtaining a machine learning prediction model. The machine learning prediction model is used to output the predicted values ​​of transportation time, node waiting time, temperature exceedance risk, and their uncertainty characterization parameters. When execution feedback data is obtained, the model parameters are updated or calibrated online based on the execution feedback data.

[0026] The parameter reasoning generation module is used to input the influenza vaccine transportation task to be planned into the machine learning prediction model, reason about the candidate transportation link set and candidate transportation node set, and obtain the transportation time prediction value, node waiting time prediction value, temperature exceedance risk prediction value and uncertainty characterization parameters corresponding to each candidate transportation link and candidate transportation node, and summarize them to form the parameter set of the transportation optimization problem.

[0027] The transport optimization solution module is used to build and solve the transport optimization model based on the parameter set, and output the influenza vaccine transport plan. The transport optimization model includes cold chain constraints, transport capacity constraints and uncertainty constraints. Cold chain constraints include temperature threshold constraints and cold preservation capacity constraints. Transport capacity constraints include time window constraints and capacity constraints. Uncertainty constraints constrain the uncertainty characterization parameters corresponding to the predicted transport time, node waiting time and temperature exceedance risk according to the preset confidence level, so that the influenza vaccine transport plan meets the time window constraints and cold chain constraints at the confidence level.

[0028] The execution feedback collection and re-optimization module is used to collect execution feedback data during the execution of the influenza vaccine transport plan. The execution feedback data includes actual transport time, actual node waiting time and temperature data during transit. Based on the execution feedback data, the module calibrates or updates the parameters of the machine learning prediction model online, and triggers the re-optimization of the influenza vaccine transport plan when the prediction deviation or constraint violation risk reaches the preset trigger condition.

[0029] This invention introduces a multi-source dynamic data-driven predictive modeling mechanism to jointly predict transportation time, node waiting time, and temperature exceedance risks during influenza vaccine transport. These uncertainties are incorporated into the transport optimization constraints and objective function, enabling the transport plan to simultaneously meet cold chain safety and timeliness requirements under a pre-set confidence level. Furthermore, by combining execution feedback data, online calibration of model parameters and dynamic re-optimization of the transport plan are achieved, effectively improving the reliability and adaptability of transport decisions in complex and variable operating environments, reducing the risks of temperature exceedances and transport failures, and improving the overall utilization efficiency of cold chain and transportation resources while ensuring vaccine quality and safety.

[0030] Example 2 This embodiment is a data-driven method and system for optimizing the transportation of influenza vaccines, illustrating the implementation process of the method of constructing training samples and training prediction models based on transportation operation data in a real business scenario.

[0031] 1. Transfer task and source of raw data This example focuses on a cold chain transportation task of influenza vaccine actually carried out by a provincial disease control and prevention center on November 18, 2022.

[0032] The starting point for this transfer mission was a provincial-level vaccine cold storage facility, and the destination was a municipal-level vaccination site, with one prefecture-level vaccine transit point in between. The transfer method was cold chain road transport, and the quantity of vaccines was 3,600 doses, loaded into 12 refrigerated boxes.

[0033] The original data sources include vehicle GPS positioning and driving logs, road congestion index provided by the provincial transportation operation platform, real-time weather data every 15 minutes provided by the meteorological department, inbound and outbound timestamps recorded by the transfer node operation system, and temperature data collected by the temperature recorder inside the cold chain box.

[0034] 2. Basis for determining sample input features Using 15 minutes before shipment, i.e. 08:15 on November 18, 2022 as the observation time, the corresponding sample input feature vector is constructed, and the determination method of each feature is as follows.

[0035] Transportation mode code: This mission will use road transportation throughout, and the corresponding coding rule for road transportation is [1,0,0].

[0036] Link path length: The planned driving distance from the starting node to the destination node is calculated to be 64.8km based on the vehicle's GPS trajectory and the navigation system's planned path.

[0037] Historical average speed and standard deviation: From 146 historical driving records of the route in the past 90 days, the average speed was 48.6 km / h and the standard deviation was 7.9 km / h.

[0038] Traffic congestion index: obtained from the real-time road operation index released by the provincial transportation operation platform at 08:15. The corresponding value is 0.38, where 0 indicates smooth traffic and 1 indicates severe congestion.

[0039] Ambient temperature and precipitation intensity: According to the meteorological department's 15-minute real-time data from the nearest meteorological station along the mission path, the ambient temperature at 08:15 was 6.2℃ and the precipitation intensity was 0mm / h.

[0040] Transfer node processing capacity per unit time: Based on the historical data of the transfer node operation system for the past 30 days, the average processing capacity of the node during normal working hours is 150 boxes / hour.

[0041] Current queue length at transit node: According to the node inbound system record at 08:15, the current number of refrigerated boxes to be processed is 22.

[0042] Historical node waiting time parameter: Based on 312 job records of the node in the past 30 days, the average waiting time of the node is 29.4 minutes.

[0043] Cold chain vehicle and refrigerant parameters: Based on the loading configuration for this mission, the rated volume of the cold chain vehicle is 2.0 m³. 3 After real-time loading, the remaining usable volume is 1.4m³. 3 .

[0044] The refrigerant weighed 15.8 kg. Based on the unit mass cold-keeping test data given in the refrigerant model manual, the theoretical cold-keeping time is calculated to be 12.6 hours.

[0045] The above data is normalized according to the system's preset feature scale parameters and then concatenated according to a predetermined feature order to form a feature vector with a dimension of 47.

[0046] 3. Methods for determining supervisory labels After the transfer task is completed, the sample output is determined based on the execution log.

[0047] Actual transport time: Calculated from the vehicle's GPS-recorded departure timestamp 08:32 and final arrival timestamp 10:03, the actual transport time is 91 minutes.

[0048] Actual node waiting time: calculated from the arrival time and handover completion time recorded by the transit node system. The actual waiting time of the transit node is 36 minutes.

[0049] Temperature exceeding limit event flag: According to the temperature recorder data inside the cold chain box, the temperature range recorded during the entire transfer process was 2.4℃-7.1℃. There were no instances where the temperature exceeded 8℃ for more than 10 minutes. Therefore, the temperature exceeding limit event flag is 0.

[0050] 4. Model Prediction Output A training sample set was constructed based on 3,260 historical transfer task samples collected within the corresponding time period. The sample was used as one of the samples in the training sample set to participate in model training and complete the model training.

[0051] After the model training is complete, the above sample inputs are substituted into the machine learning prediction model. The machine learning prediction model can be implemented using a probabilistic prediction model, and the output results are as follows: The predicted location parameter for transportation time is 88.7 min, the transportation time scale parameter is 6.1 min, the predicted location parameter for node waiting time is 33.2 min, the node waiting time scale parameter is 4.5 min, the predicted probability of temperature exceeding the limit is 0.06, and the risk confidence interval at a 95% confidence level is [0.02, 0.11].

[0052] The deviation between the above prediction results and the actual execution results did not exceed the preset thresholds for the mean deviation of the prediction residuals and the threshold for the increase in residual variance, so the online update of the model parameters was not triggered.

[0053] The predicted location parameters, scale parameters, and temperature exceedance risk confidence interval parameters of the transportation time output by the prediction model are used as uncertainty input parameters for time window constraints and cold chain constraints in the subsequent transshipment optimization model.

[0054] Through the above implementation steps, feature vectors are constructed based on vehicle driving data, node operation data, meteorological data, and cold chain status data collected from actual transfer tasks. These feature vectors are then used to form training samples for model training. The resulting machine learning prediction model can predict transportation time, node waiting time, and the risk of temperature exceeding limits, and output corresponding uncertainty characterization parameters. In this embodiment's transfer task, the deviation between the model prediction results and the actual execution results is within a preset range and does not meet the triggering conditions for online model parameter updates, indicating that the method operates stably in the actual influenza vaccine transfer scenario and provides reliable parameter input for subsequent transfer optimization models.

[0055] Example 3 This embodiment describes the implementation of the influenza vaccine transport management system in actual operation, based on a real-world influenza vaccine transport scenario, including the construction and solution of the transport optimization model and the generation of transport plans.

[0056] 1. System Composition and Entity Performing Transfer Tasks In October 2024, a provincial-level Center for Disease Control and Prevention (CDC) conducted an influenza vaccine transport mission. The transport was completed using an information system deployed within the provincial CDC. This information system included: The data acquisition server is used to access data from transportation platforms, cold chain equipment, and node operation systems.

[0057] Predictive computing servers are used to predict transportation time, node waiting time, and temperature risks.

[0058] The transportation optimization solution module is used to generate transportation schemes under constraints.

[0059] Cold chain transport vehicles and on-the-road temperature monitoring terminals are used to execute and report data during the transfer process.

[0060] All of the above systems are online and are uniformly scheduled by the same control system in this implementation.

[0061] 2. Acquisition and processing of real business data On October 12, 2024, the system received a vaccine transfer instruction, which stated that 16,000 doses of influenza vaccine should be transferred from the provincial cold chain warehouse to the centralized vaccination site in City A.

[0062] The data acquisition server automatically pulls data from the following systems: Historical transit database: We retrieved 37 historical transport records for the same type of vaccine via the same route from September 2023 to September 2024. Each record includes the actual transport time, temperature curve during transit, and transfer operation time.

[0063] Traffic operation data interface: Obtain the average vehicle speed and congestion index of highways and ordinary roads during the period from 06:00 to 08:00 on the day of transfer.

[0064] Cold chain vehicle management system: obtains the rated loading capacity of available cold chain vehicles (rated loading capacity is 20,000 doses), historical cold preservation test data, and refrigerant loading records.

[0065] Transit node operation system: Obtain statistical data on queuing and processing time of municipal cold chain transit warehouses over the past 12 months.

[0066] After the data acquisition server completes time alignment of the above data, it sends the processing results to the prediction calculation server.

[0067] 3. Formation of prediction parameters The predictive computing server is a server device equipped with a processor and memory. The memory stores program instructions used to perform statistical modeling and probabilistic prediction. Based on sample data in a historical transit database, the predictive computing server filters samples that match the time period of the current task.

[0068] For the high-speed direct route, the server selected 22 historical transit records executed between 07:00 and 09:00, and calculated the following: The average actual transport time was 6.18 hours, and the standard deviation of the actual transport time was 0.79 hours.

[0069] Based on the above statistical results, the prediction calculation server estimates the transportation time distribution parameters through the built-in probability prediction model, and outputs the transportation time prediction parameters accordingly: the predicted mean transportation time is 6.2 hours and the scale parameter is 0.8 hours.

[0070] For the municipal-level cold chain transit warehouse node, the server selected 19 transit records with the same time period from the historical data, obtained the average node waiting time of 0.92 hours and the standard deviation of 0.31 hours, and output the node waiting time prediction parameters as 0.9 hours / 0.3 hours.

[0071] The above sample size reaches the system's preset sample size threshold for statistical modeling of similar transfer tasks, satisfying the stability requirements of model parameter estimation.

[0072] 4. Calculation and generation of temperature exceedance risk The predictive computing server performed statistical analysis on the in-transit temperature curves of the above 22 historical transport records, among which: There were 4 instances of temperatures exceeding 8°C and lasting for more than 5 minutes. After verification by the cold chain equipment management system, 3 of these instances corresponded to insufficient refrigerant loading or abnormal equipment operation.

[0073] When calculating risk parameters, the server automatically removed abnormal operating condition samples and only kept 19 records of normal equipment status, resulting in: 1 temperature over-limit event occurred, and the empirical risk probability was approximately 1.8%.

[0074] Based on the sample size, historical risk distribution characteristics, and binomial event confidence interval estimation method, the prediction calculation server calculates the upper bound of risk at a 95% confidence level as 3.0%, and sends the result as a cold chain risk constraint parameter to the transshipment optimization solution module.

[0075] 5. Execution subject and solution process of the transportation optimization model After receiving the above prediction parameters, the transit optimization solution module constructs a mixed-integer optimization model that includes binary variables for path selection, and sets the following constraints: The total transfer time limit is set based on the vaccine cold chain management regulations and the statistical results of the execution time of similar vaccine transfer tasks in the past, and is set to no more than 12 hours, with the cold chain vehicle loading capacity not less than 16,000 doses.

[0076] Historical cold chain vehicle cold preservation test data shows that, under the condition of configuring the corresponding refrigerant, its continuous cold preservation capacity is greater than the upper limit of the total transfer time.

[0077] The confidence level is pre-configured by the influenza vaccine transport management system to balance transport timeliness and cold chain safety risks in transport planning. At a 95% confidence level, the upper bound of the quantiles of transport time and node waiting time must meet time window constraints, and the upper bound of the temperature exceedance risk must not exceed 5%.

[0078] The upper bound of the transit time quantile is calculated by the solution module in the following way, using the formula:

[0079] in, This represents the corresponding uncertainty scale parameter. Indicates the preset credit level. To be consistent with confidence level The corresponding quantile coefficient is 1.645, which is used by the system to approximate the transportation time and waiting time as a normal distribution.

[0080]

[0081]

[0082] The sum of the two results in an upper limit of 8.91 hours for the total transit time.

[0083] Under the premise of satisfying all constraints, the transfer optimization solution module completes the solution with the weighted sum of transportation time and temperature risk as the objective, and determines that the high-speed direct route is adopted as the transfer plan for this project. The weight parameters corresponding to transportation time and temperature risk are automatically determined by the system based on the vaccine type and cold chain safety level, through the weight configuration rules pre-stored in the system.

[0084] 6. Implementation and Results Feedback Based on the solution results, the system control unit issued a dispatch instruction to the cold chain vehicle at 08:30 on October 12, 2024.

[0085] During the transfer process, the temperature monitoring terminal collected temperature data every minute and transmitted it back to the system in real time. After the transfer was completed, the system recorded: The actual transport time was 7.1 hours, the actual transfer waiting time was 1.0 hour, the temperature range throughout the journey was 2.6-7.4℃, and no temperature exceeding the limit occurred.

[0086] The system writes the above execution feedback data into the historical transfer database for parameter updates in subsequent transfer tasks.

[0087] In summary, during the actual execution of the transport plan generated by the embodiment, the actual transport time and node waiting time did not exceed the upper limit of the time determined under the preset information level, and the temperature during transit was always within the safe range specified for vaccines, with no temperature exceeding the limit. Therefore, it can be seen that in this real transport task, the transport plan generated by the influenza vaccine transport management system met the time window constraints and cold chain safety constraints during actual execution, and the system operation results were consistent with the transport planning results.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data-driven optimization method for influenza vaccine transport, characterized in that, include: S1. Obtain multi-source dynamic data corresponding to the influenza vaccine transport task, and perform time alignment and feature construction on the multi-source dynamic data to obtain feature vectors for characterizing the candidate transport link set and the candidate transport node set. The candidate transport link is used to characterize the transport mode and path combination from the starting node to the destination node, and the candidate transport node is used to characterize the operation and waiting process of loading, unloading, handover or transfer nodes. S2. Construct a training sample set based on the feature vector, and perform model training on the training sample set to learn model parameters, thereby obtaining a machine learning prediction model. The machine learning prediction model is used to output the predicted value of transportation time, the predicted value of node waiting time, the predicted value of temperature exceeding the limit risk and its uncertainty characterization parameters. When the execution feedback data is obtained, the model parameters are updated or calibrated online based on the execution feedback data. S3. Input the influenza vaccine transport task to be planned into the machine learning prediction model, and reason about the candidate transport link set and the candidate transport node set to obtain the transport time prediction value, node waiting time prediction value, temperature exceedance risk prediction value and uncertainty characterization parameters corresponding to each candidate transport link and each candidate transport node, and summarize them to form the parameter set of the transport optimization problem. S4. Based on the parameter set, construct and solve the transfer optimization model to output the influenza vaccine transfer plan. The transfer optimization model includes cold chain constraints, transportation capacity constraints, and uncertainty constraints. The cold chain constraints include temperature threshold constraints and cold preservation capacity constraints. The transportation capacity constraints include time window constraints and capacity constraints. The uncertainty constraints constrain the uncertainty characterization parameters corresponding to the predicted transportation time, predicted node waiting time, and predicted temperature exceedance risk according to a preset confidence level, so that the influenza vaccine transfer plan satisfies the time window constraints and the cold chain constraints at the confidence level. The influenza vaccine transfer plan includes the selection results of candidate transfer links and candidate transfer nodes, the departure time and node operation period, and the vehicle and refrigerant configuration corresponding to the cold preservation capacity constraints. S5. During the execution of the influenza vaccine transport plan, execution feedback data is collected, including actual transport time, actual node waiting time and temperature data during transit. The parameters of the machine learning prediction model are calibrated or updated online based on the execution feedback data. When the prediction deviation or constraint violation risk reaches the preset trigger condition, the influenza vaccine transport plan is re-optimized.

2. The data-driven influenza vaccine transport optimization method according to claim 1, characterized in that: The multi-source dynamic data includes road traffic operation data, meteorological data, operational capacity data and queuing data of each candidate transfer node, availability and capacity data of cold chain vehicles, refrigerant inventory and replenishment data, and historical temperature data in transit. The multi-source dynamic data is uniformly mapped to a time series based on a preset time granularity before the time alignment is performed.

3. The data-driven influenza vaccine transport optimization method according to claim 1, characterized in that: The feature construction includes: constructing link-level features for the candidate transfer links and node-level features for the candidate transfer nodes; concatenating the link-level features and node-level features in a preset feature order to generate the feature vector; the link-level features include transportation mode codes, path length, historical speed distribution parameters, meteorological impact features, and traffic congestion index; the node-level features include node operation period codes, unit time processing capacity, queue length statistics, historical waiting time distribution parameters, and node temperature control environment features.

4. The data-driven influenza vaccine transport optimization method according to claim 1, characterized in that: The training sample set is constructed from historical transfer tasks. Each training sample in the training sample set consists of a sample input and a sample output. The sample input is the feature vector extracted for the same historical transfer task within a preset observation time window. The sample output is a supervision label determined by the historical records of the historical transfer task on the candidate transfer link or candidate transfer node corresponding to the feature vector. The supervision label includes the actual transport time, the actual node waiting time, and the temperature exceedance event marker.

5. The data-driven influenza vaccine transport optimization method according to claim 1, characterized in that: The machine learning prediction model is a probabilistic prediction model. The machine learning prediction model outputs location parameters and scale parameters as uncertainty characterization parameters for the predicted value of transportation time and the predicted value of node waiting time, and outputs risk probability value and risk confidence interval parameters as uncertainty characterization parameters for the predicted value of temperature exceeding limit risk.

6. The data-driven influenza vaccine transport optimization method according to claim 1, characterized in that: The online update or calibration includes: calculating the predicted residual based on the execution feedback data, and performing incremental learning update on the model parameters when the statistics of the predicted residual meet the preset drift criterion, wherein the drift criterion is a residual mean offset threshold or a residual variance increase threshold.

7. The data-driven influenza vaccine transport optimization method according to claim 1, characterized in that: The transfer optimization model is a mixed integer optimization model that includes binary decision variables. The binary decision variables are used to represent the selection results of the candidate transfer links and the candidate transfer nodes. The objective function of the transfer optimization model includes a transport time objective item and a temperature exceedance risk objective item. The objective function performs a weighted summation of each objective item using preset weights.

8. The data-driven influenza vaccine transport optimization method according to claim 1, characterized in that: The uncertainty constraint is implemented through an opportunity constraint. Under the preset confidence level, the predicted transport time and the corresponding uncertainty characterization parameters are converted into corresponding quantile upper bounds, and the predicted node waiting time and the corresponding uncertainty characterization parameters are converted into corresponding quantile upper bounds. The quantile upper bounds are then used to replace the deterministic duration parameters in the time window constraint, so that the influenza vaccine transport scheme satisfies the time window constraint under the confidence level.

9. The data-driven influenza vaccine transport optimization method according to claim 1, characterized in that: The preset triggering conditions include: the deviation between the actual transportation time and the predicted transportation time exceeds a first threshold; the deviation between the actual node waiting time and the predicted node waiting time exceeds a second threshold; the duration for which the temperature data in transit exceeds the temperature threshold constraint reaches a third threshold; and the predicted value of temperature exceedance risk obtained based on real-time inference exceeds a fourth threshold. When the re-optimization is triggered, the candidate transfer links and candidate transfer nodes that have been executed are locked as fixed decisions, and the transfer optimization model is re-solved for the unexecuted parts to update the departure time, node operation period, and vehicle and refrigerant configuration.

10. A data-driven influenza vaccine transport optimization system, implemented according to the data-driven influenza vaccine transport optimization method described in claim 9, characterized in that, include: Feature construction module: used to acquire multi-source dynamic data corresponding to the influenza vaccine transport task, and to perform time alignment and feature construction on the multi-source dynamic data to obtain feature vectors that characterize the candidate transport link set and the candidate transport node set; Prediction model module: used to construct a training sample set based on the feature vector, and perform model training on the training sample set to learn model parameters, thereby obtaining a machine learning prediction model. The machine learning prediction model is used to output the predicted value of transportation time, the predicted value of node waiting time, the predicted value of temperature exceedance risk and its uncertainty characterization parameters, and when the execution feedback data is obtained, the model parameters are updated or calibrated online based on the execution feedback data. Parameter reasoning generation module: It is used to input the influenza vaccine transportation task to be planned into the machine learning prediction model, reason about the candidate transportation link set and the candidate transportation node set, and obtain the transportation time prediction value, node waiting time prediction value, temperature exceedance risk prediction value and uncertainty characterization parameters corresponding to each candidate transportation link and each candidate transportation node, and summarize them to form the parameter set of the transportation optimization problem. The transfer optimization solution module is used to construct and solve the transfer optimization model based on the parameter set, and output the influenza vaccine transfer plan. The transfer optimization model includes cold chain constraints, transportation capacity constraints, and uncertainty constraints. The cold chain constraints include temperature threshold constraints and cold preservation capacity constraints. The transportation capacity constraints include time window constraints and capacity constraints. The uncertainty constraints constrain the uncertainty characterization parameters corresponding to the predicted values ​​of transportation time, node waiting time, and temperature exceedance risk according to a preset confidence level, so that the influenza vaccine transfer plan satisfies the time window constraints and the cold chain constraints at the confidence level. The execution feedback collection and re-optimization module is used to collect execution feedback data during the execution of the influenza vaccine transport plan. The execution feedback data includes actual transport time, actual node waiting time and temperature data during transit. Based on the execution feedback data, the module calibrates or updates the parameters of the machine learning prediction model online, and triggers the re-optimization of the influenza vaccine transport plan when the prediction deviation or constraint violation risk reaches a preset trigger condition.

Images