A dynamic cold chain distribution strategy optimization method and system

Abstract

The application provides a dynamic cold chain distribution strategy optimization method and system, belonging to the technical field of fresh food cold chain distribution intelligent scheduling optimization, which comprises the following steps: based on historical and real-time order data analysis, determining a service coverage range and generating a benchmark distribution path, which together constitute a time-space constraint region; obtaining road condition data, environmental temperature data and vehicle operation data, obtaining a first disturbance feature vector and inputting a distribution state evaluation model to obtain a stability evaluation result and a disturbance level classification, obtaining a current distribution path, determining a target optimization path segment, generating a reconstructed path segment and splicing and extracting it with the unadjusted part of the path to generate a second distribution path feature vector, inputting the second distribution path feature vector and the benchmark distribution path feature vector into a path deviation identification model to obtain a risk identification result, combining the first disturbance feature vector analysis, extracting multi-dimensional disturbance feature variables and inputting them into a path planning and distribution scheduling model with the benchmark distribution path feature vector to obtain a final distribution path scheme.

Description

Technical Field

[0001] This application relates to the technical field of intelligent scheduling and optimization of fresh food cold chain distribution, and in particular to a dynamic cold chain distribution strategy optimization method and system. Background Technology

[0002] Fresh food cold chain delivery has stringent requirements for timeliness and temperature control. Existing route planning relies on static baseline routes generated from historical orders and manual adjustments, which has its limitations. Current solutions cannot integrate multi-source dynamic disturbance data such as real-time road conditions, ambient temperature, and vehicle range, making it impossible to quantitatively assess the stability of the delivery network or accurately identify high-risk sections. Furthermore, the lack of a standardized vector extraction mechanism prevents comparison of deviations and delay risks between adjusted routes and the initial baseline route, resulting in a disconnect between disturbance factors and route optimization. Existing optimization algorithms use fixed parameters for optimization, failing to adaptively adjust the search range based on real-time disturbance levels. They only output the planned route once, lacking a complete closed loop of disturbance identification, local reconstruction, deviation verification, and global scheduling. When sudden disturbances such as congestion or high temperatures occur, route adjustments lag behind, leading to prolonged exposure and temperature loss of fresh produce, increasing spoilage rates. Current technologies do not delineate spatiotemporal constraint areas for unified vehicle service range management, preventing deep integration of multi-dimensional disturbance features for global scheduling. Route optimization objectives cannot simultaneously consider delivery time, fresh produce deterioration, and driving reliability, resulting in insufficient overall adaptive scheduling capabilities.

[0003] Therefore, there is an urgent need for a dynamic cold chain distribution strategy optimization method and system. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a method and system for optimizing dynamic cold chain distribution strategies.

[0005] A first aspect of this application provides a method for optimizing a dynamic cold chain distribution strategy, comprising: Based on historical and real-time order data of the target delivery area, the service coverage of each cold chain delivery vehicle is determined by analyzing the spatiotemporal distribution of cold chain delivery vehicles and the quality degradation characteristics of fresh goods, and an initial delivery route is generated as the benchmark delivery route; the service coverage and the benchmark delivery route together constitute the first spatiotemporal constraint area. Feature extraction is performed on the baseline delivery route to obtain the baseline delivery route feature vector; Within the first spatiotemporal constraint region, road condition data, ambient temperature data, and vehicle operation data are acquired, and the road condition data, ambient temperature data, and vehicle operation data are preprocessed and feature extracted, and then fused to obtain the first disturbance feature vector. The first disturbance feature vector is input into a preset delivery status assessment model to obtain the stability assessment result and disturbance level classification of the current delivery network. Obtain the current delivery route of the cold chain delivery vehicle, and based on the stability assessment results and disturbance level classification, determine the target optimized route segment that needs to be replanned or partially adjusted from the current delivery route; The target optimized path segment is partially reconstructed and adjusted to generate a reconstructed path segment. The reconstructed path segment is then spliced ​​with the unadjusted part of the current delivery path and its features are extracted to generate a second delivery path feature vector. The second delivery path feature vector and the baseline delivery path feature vector are then input into a preset path deviation identification model to obtain the path deviation and delay risk identification results. The path offset and delay risk identification results and the first disturbance feature vector are fused and analyzed to extract multi-dimensional disturbance feature variables; The multi-dimensional perturbation feature variables and the baseline delivery path feature vector are input into a preset adaptive path planning and delivery scheduling model to obtain the final delivery path scheme after optimization.

[0006] A second aspect of this application provides a dynamic cold chain distribution strategy optimization system, comprising: The region determination module is used to determine the service coverage of each cold chain delivery vehicle based on historical and real-time order data of the target delivery area by analyzing the spatiotemporal distribution of cold chain delivery vehicles and the quality degradation characteristics of fresh goods, and to generate an initial delivery route as a benchmark delivery route; the service coverage and the benchmark delivery route together constitute the first spatiotemporal constraint region. The feature extraction module is used to extract features from the baseline delivery route to obtain a feature vector of the baseline delivery route; The disturbance vector module is used to acquire road condition data, ambient temperature data and vehicle operation data within the first spatiotemporal constraint region, and to preprocess and extract features from the road condition data, ambient temperature data and vehicle operation data, and then fuse them to obtain the first disturbance feature vector. The status assessment module is used to input the first disturbance feature vector into a preset delivery status assessment model to obtain the stability assessment result and disturbance level classification of the current delivery network. The target optimization module is used to obtain the current delivery route of the cold chain delivery vehicle, and based on the stability assessment results and disturbance level classification, determine the target optimization route segment that needs to be replanned or partially adjusted from the current delivery route. The path reconstruction module is used to partially reconstruct and adjust the target optimized path segment to generate a reconstructed path segment. The reconstructed path segment is then spliced ​​with the unadjusted part of the current delivery path and its features are extracted to generate a second delivery path feature vector. The second delivery path feature vector and the baseline delivery path feature vector are then input into a preset path deviation identification model to obtain the path deviation and delay risk identification results. The multidimensional vector module is used to fuse and analyze the path offset and delay risk identification results and the first disturbance feature vector to extract multidimensional disturbance feature variables. The optimization module is used to input the multi-dimensional perturbation feature variables and the baseline delivery path feature vector into a preset adaptive path planning and delivery scheduling model to obtain the final delivery path scheme after optimization.

[0007] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described dynamic cold chain distribution strategy optimization method.

[0008] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described dynamic cold chain distribution strategy optimization method.

[0009] The beneficial effects of the dynamic cold chain delivery strategy optimization method and system provided in this application are as follows: This application delineates spatiotemporal constraint areas based on orders, vehicles, and fresh produce deterioration characteristics; constructs disturbance feature vectors using multi-source real-time data; quantifies the disturbance level of the delivery network using an evaluation model; accurately locates local reconstruction of high-risk road segments; and avoids computational redundancy caused by global replanning. It integrates path deviation risk and multi-dimensional disturbance variables into an adaptive scheduling model, simultaneously considering delivery time, fresh produce preservation, and path reliability. It achieves dynamic local path optimization driven by disturbance hierarchy, accurately predicts path deviation and delay risks, and reduces fresh produce quality loss; based on real-time road conditions and cargo status, it adaptively adjusts the optimization strategy, improves delivery network stability, shortens delivery delay time, balances scheduling computation efficiency and fresh produce delivery quality, and enables large-scale multi-vehicle real-time cold chain scheduling. Attached Figure Description

[0010] Figure 1 A flowchart illustrating a dynamic cold chain distribution strategy optimization method provided in an embodiment of this application; Figure 2 This is a structural block diagram of a dynamic cold chain distribution strategy optimization system provided in an embodiment of this application; Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0012] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1 - Appendix Figure 3 The following is an explanation using specific examples.

[0013] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a dynamic cold chain distribution strategy optimization method provided in an embodiment of this application. The method includes: S101: Based on historical and real-time order data of the target delivery area, the service coverage of each cold chain delivery vehicle is determined by analyzing the spatiotemporal distribution of cold chain delivery vehicles and the quality degradation characteristics of fresh goods, and an initial delivery route is generated as the benchmark delivery route; the service coverage and the benchmark delivery route together constitute the first spatiotemporal constraint area.

[0014] In this embodiment, the target delivery area is a designated geographical region for fresh food cold chain delivery, which can be divided according to administrative districts, business districts, and logistics transit stations. It serves as the geographical boundary for route optimization operations, and all order and vehicle data collection and road condition monitoring are limited to this area. Historical order data consists of all orders completed within the target delivery area over a fixed period in the past, including order address, product category, order time, delivery completion time, fresh food spoilage records, customer time windows, vehicle allocation records, etc., used to discover delivery patterns.

[0015] In this embodiment, real-time order data refers to the data of newly added orders to be delivered at the current time of scheduling, including real-time delivery address, fresh produce category, required delivery time, initial temperature of goods, and order weight and volume. This data serves as the immediate basis for dynamically adjusting the service range of vehicles. The spatiotemporal distribution of cold chain delivery vehicles refers to the distribution patterns of multiple cold chain vehicles in the target area at different times, including location, capacity, frequency of travel, frequently used routes, and empty / full load time. The spatial dimension reflects the vehicle's permanent service area, while the temporal dimension represents the morning, noon, and evening delivery peaks and idle periods.

[0016] In this embodiment, the quality degradation characteristics of fresh produce refer to the quality loss patterns of different fresh produce categories as transportation time and ambient temperature change, including degradation rate, critical shelf life, and temperature threshold. For example, leafy greens rapidly lose water and spoil at high temperatures, and frozen meats spoil faster after temperature fluctuations. This is used to constrain the maximum allowable delivery time. The service coverage area is the geographical region exclusively responsible for by a single cold chain delivery vehicle. All orders within the region are preferentially assigned to that vehicle. The scope of coverage takes into account order density, vehicle load limit, and maximum shelf life of fresh produce to avoid aggravating losses due to long-distance transportation across regions.

[0017] In this embodiment, the initial delivery route / baseline delivery route is a static standard delivery route calculated solely based on historical order patterns, static road network, and standard shelf-life, without incorporating dynamic disturbances such as real-time congestion or temporary temperature increases. It serves as a benchmark for subsequent dynamic optimization, used to compare route deviations and loss changes after adjustment. The first spatiotemporal constraint region is a set of boundaries that integrates both spatial and temporal constraints. Spatially, it restricts each vehicle to deploy delivery nodes only within its own service coverage area; temporally, it follows the arrival and stopping time intervals of each node preset in the baseline route. Subsequent road condition, temperature, vehicle data collection, and route reconstruction operations cannot exceed the boundaries of this region.

[0018] S102: Extract features from the baseline delivery route to obtain the feature vector of the baseline delivery route.

[0019] In this embodiment, the baseline delivery route is a generated static standard delivery route that does not include dynamic disturbances such as real-time congestion and temperature fluctuations. It includes complete planning information such as vehicle passing nodes, node sequence, planned arrival / stop time, mileage, load, and preset temperature control range throughout the route, and serves as a standard reference for comparing the merits of different routes.

[0020] In this embodiment, feature extraction is a process of filtering quantitative indicators strongly correlated with cold chain optimization from complete delivery route information, eliminating invalid and redundant information, and converting unstructured data such as text, geography, and time into calculable numerical values. Extraction dimensions include four categories: spatial, temporal, fresh produce preservation, and vehicle capacity. The baseline delivery route feature vector is a one-dimensional numerical array formed by arranging all standardized feature values ​​extracted from the baseline route in a fixed order. Each value corresponds to a path feature, and the dimension is fixed. The entire route serves as a comparison benchmark, used for difference calculation, risk identification, and model input with the adjusted route feature vector. Common feature dimensions include: total mileage, number of nodes passed through, planned arrival time of each node, average duration of a single road segment, maximum allowable decay of fresh produce throughout the entire route, vehicle rated remaining volume, preset ambient temperature thresholds for each road segment, total planned delivery time, percentage of peak road segments, and customer time window tightness.

[0021] S103: Within the first spatiotemporal constraint region, acquire road condition data, ambient temperature data, and vehicle operation data, and preprocess and extract features from the road condition data, ambient temperature data, and vehicle operation data, and obtain the first disturbance feature vector after fusion.

[0022] In this embodiment, road condition data refers to real-time traffic information within the constrained area, including road congestion levels, real-time vehicle speeds, road construction, traffic light waiting times, traffic volume, and road restriction information. This is an external source of traffic disturbance, directly affecting delivery travel time and fresh produce exposure time. Ambient temperature data includes real-time outdoor temperature, sunlight intensity, and ventilation conditions along the delivery route. Ambient temperature accelerates heat exchange within the refrigerated truck compartment, increasing the rate of fresh produce quality deterioration; it is a unique disturbance variable specific to the cold chain. Vehicle operation data consists of real-time operating parameters of the cold chain delivery vehicle itself, including actual compartment cooling temperature, door opening and closing frequency, remaining battery / fuel level, current cargo weight, instantaneous speed, and braking / stopping frequency; representing the vehicle's own cooling and transport capacity disturbances.

[0023] In this embodiment, preprocessing involves cleaning, correcting, supplementing, and normalizing the raw collected data: removing abnormal and garbled values ​​from sensors, filling in missing data due to signal interruptions, eliminating dimensional differences, and removing redundant data outside the constraint region to ensure data validity and usability. Feature extraction involves selecting quantitative indicators strongly correlated with delivery disturbances from the three types of raw data, transforming text, time series, and raw sensor signals into standardized numerical features, and distinguishing between three feature dimensions: traffic disturbances, temperature disturbances, and vehicle-specific disturbances.

[0024] In this embodiment, fusion involves concatenating features extracted from three types of data—road conditions, ambient temperature, and vehicle operation—in a predetermined fixed order, unifying dimensions and numerical ranges, and integrating them into an ordered feature set. The first perturbation feature vector is a fixed-dimensional one-dimensional numerical array obtained after fusion, with each component corresponding to a real-time perturbation feature. This quantifies the degree of all dynamic disturbances in the current delivery scenario and serves as the input parameter for the delivery status assessment model.

[0025] S104: Input the first disturbance feature vector into the preset delivery status assessment model to obtain the stability assessment result and disturbance level classification of the current delivery network.

[0026] In this embodiment, the preset delivery status evaluation model is a multi-input evaluation model that has been trained offline in advance. This embodiment uses a lightweight neural network model.

[0027] Specifically, the preset delivery status assessment model is a lightweight neural network model with three layers of data flow: feature preprocessing layer, lightweight feature extraction layer, and dual-branch output layer. The first perturbation feature vector completes standardization, feature mining, regression, and classification prediction layer by layer. Low-parameter operators are used throughout the process to adapt to the real-time operation of the vehicle terminal.

[0028] The pre-defined delivery status assessment model is built using a lightweight one-dimensional deep separable convolution. The first preprocessing layer normalizes the first perturbation feature vector to eliminate dimensional differences. The second layer extracts perturbation-related features such as road conditions, temperature, and timeliness through a small number of 1×1 convolutional kernels, and compresses the dimension using global average pooling. The third layer splits the regression and classification into a fully connected network, which respectively fits the stability score and distinguishes the perturbation category, and incorporates lightweight Dropout to suppress overfitting.

[0029] The input to the preset delivery status assessment model is a one-dimensional first perturbation feature vector that integrates multi-dimensional delivery interference indicators. The model has two outputs: one is the delivery network stability assessment result with continuous values ​​in the 0-1 range, where higher values ​​indicate more stable operation; the other is a discrete perturbation level classification, which outputs only two labels: slight perturbation and severe perturbation. The output results are used to adaptively adjust the hyperparameters of the particle swarm optimization.

[0030] In this embodiment, the current delivery network refers to the local delivery sub-network within the first spatiotemporal constraint area corresponding to a single cold chain vehicle, which is composed of the base path, each delivery node, the route segment, and the on-board cold chain equipment, and is different from the total delivery network of all vehicles in the entire region.

[0031] In this embodiment, the stability assessment result is a quantitative score output by a preset delivery status assessment model. A higher score indicates less overall disturbance in the current road network, temperature control, and vehicle operating conditions, resulting in a more stable delivery process. Conversely, a lower score indicates more severe disturbance, leading to higher risks of fresh produce spoilage and delivery delays. The disturbance level classification is a discrete grading label based on the stability score range, used by the preset delivery status assessment model to qualitatively distinguish disturbance intensity. In this embodiment, it is divided into two levels: minor disturbance and severe disturbance. Different levels correspond to subsequent path adjustment strategies and differentiated configuration rules for particle swarm optimization hyperparameters.

[0032] S105: Obtain the current delivery route of the cold chain delivery vehicle, and based on the stability assessment results and disturbance level classification, determine the target optimized route segment that needs to be replanned or partially adjusted from the current delivery route.

[0033] In this embodiment, the current delivery route of the cold chain delivery vehicle is a dynamic delivery route that is being executed in real time and transmitted back by the vehicle's on-board terminal. This route includes completed sections, untraveled sections, all receiving and stopping points along the way, and the planned arrival and stopping times at each point. It is continuously updated as the vehicle's journey progresses, which is different from the fixed baseline delivery route.

[0034] In this embodiment, route replanning is a global optimization method that reorders all delivery stops and replans the entire route. This method has high computational overhead and is only used when the risk of the entire route generally exceeds the standard. Local adjustment is a local optimization method that only replaces the roads within a certain high-risk section. The start and end stops of this section, the other delivery routes before and after it, and the stop order are all retained. This method has low computational consumption and is the preferred optimization method used in this embodiment.

[0035] In this embodiment, the target optimized path segment is a continuous road segment interval with high overall disturbance risk on the current delivery route, and is the object of the path reconstruction operation; it includes the corrected starting stop node, the ending stop node, and the original high-risk road set between the two points.

[0036] S106: Perform local reconstruction and adjustment on the target optimized path segment to generate a reconstructed path segment. Then, concatenate the reconstructed path segment with the unadjusted part of the current delivery path and extract its features to generate a second delivery path feature vector. Finally, input the second delivery path feature vector and the baseline delivery path feature vector into a preset path deviation identification model to obtain the path deviation and delay risk identification results.

[0037] In this embodiment, the target optimized path segment is the output continuous high-risk road segment with definite start and end points. It is the only section that needs to be replanned, while the rest of the delivery segments remain unchanged. Local reconstruction and adjustment does not change the start and end delivery points of the road segment, but only changes the passing roads within the road network between the two points and adjusts the branch roads within the road segment. It takes into account multiple objectives such as delivery time, fresh food preservation, and road reliability for optimization, without changing the stop order of the entire delivery order, which is different from global replanning.

[0038] In this embodiment, the reconstructed path segment is a low-risk alternative path segment recalculated using a local optimization algorithm, with the start and end nodes remaining unchanged and the internal access roads updated. The unadjusted portion of the current delivery route is all the original delivery route excluding the target optimized path segment, including already driven sections, low-risk sections waiting to be traversed, and the order of all stops. The original route and time plan are completely preserved during the splicing process.

[0039] In this embodiment, splicing involves replacing the target optimized path segment in the original path with the reconstructed new path segment, connecting the unadjusted path segments, and combining them to generate a complete, continuous, and entirely new delivery route with the delivery node order remaining unchanged. Feature extraction involves extracting quantitative indicators such as mileage, duration, temperature control, and time window from the spliced ​​and updated complete delivery route.

[0040] In this embodiment, the second delivery path feature vector is a standardized one-dimensional numerical vector obtained by extracting all features from the updated complete delivery route and arranging them in a fixed order. Its dimension and feature order are completely consistent with the baseline delivery path feature vector, and it is used for horizontal comparison of deviations. The baseline delivery path feature vector is a feature vector extracted based on a static, undisturbed standard route, serving as a standard reference for determining the quality of the path and the magnitude of the offset.

[0041] In this embodiment, the preset path deviation identification model is a pre-trained quantitative identification model. It takes two sets of path feature vectors of the same dimension as input, automatically calculates the differences in various features between the two routes, and outputs quantitative results of deviation and delay-related risks. The preset path deviation identification model in this embodiment uses a gradient boosting regression tree model. The path deviation and delay risk identification results are comprehensive judgment data output by the model, including quantitative indicators such as spatial route deviation, total delivery time increment, additional attenuation of fresh produce, and timeout probability of each node's time window, characterizing the deviation risk of the optimized route relative to the standard baseline route.

[0042] Specifically, the preset path deviation identification model adopts a gradient boosting regression tree model, which goes through three levels: feature fusion layer, multi-decision tree hierarchical iterative operation layer, and result integration output layer. After the feature vector of the second delivery path is fused with the feature vector of the benchmark delivery path, feature splitting, single-tree deviation fitting, and multi-tree result weighted integration are completed layer by layer, and finally the comprehensive identification result is output.

[0043] The preset path deviation identification model is built on the gradient boosting regression tree as the basic framework. The feature fusion layer concatenates and integrates the two sets of path feature vectors to eliminate dimensional differences. The multi-layer decision tree sequentially divides and fits the trajectory deviation and duration deviation features, and each tree iteratively corrects the previous residual. The ensemble layer weightedly summarizes the outputs of all decision trees. During training, the planned path is paired with actual execution samples, and hyperparameters such as learning rate and tree depth are iteratively tuned. After convergence, the model is solidified for real-time identification.

[0044] The preset path deviation identification model takes two sets of one-dimensional feature vectors as input: the feature vector of the second delivery path and the feature vector of the baseline delivery path. The model outputs the path deviation and delay risk identification results, which consist of the path deviation quantification value and the delay risk level. These results are used to generate a plan execution deviation report and support the closed-loop correction of scheduling model parameters.

[0045] S107: The path deviation and delay risk identification results and the first disturbance feature vector are fused and analyzed to extract multi-dimensional disturbance feature variables.

[0046] In this embodiment, the fusion analysis unifies the dimension and dimensional mapping relationship between the path deviation and delay risk identification results and the first disturbance feature vector, establishes the correlation and coupling relationship between the original disturbance and the path deviation and delay risk, eliminates redundant and duplicate features, completes the joint weighted correlation operation of the two types of data, and realizes the linkage analysis of static path deviation risk and real-time dynamic disturbance information.

[0047] In this embodiment, the multi-dimensional disturbance feature variables are a set of independent quantitative indicators extracted after fusion analysis. They include the original environmental traffic disturbance dimension, the offset delay risk dimension caused by path adjustment, and take into account the real-time on-site interference and the secondary loss and timeliness risk after path reconstruction. It is a set of multiple variables used as input to the adaptive path planning model.

[0048] S108: Input the multi-dimensional perturbation feature variables and the baseline delivery path feature vector into the preset adaptive path planning and delivery scheduling model to obtain the final delivery path scheme after optimization.

[0049] In this embodiment, the preset adaptive path planning and delivery scheduling model optimization objectives include minimizing delivery time, minimizing fresh produce quality degradation, maximizing path reliability, and outputting a compliant and complete delivery route. This embodiment adopts a multi-objective adaptive genetic algorithm model.

[0050] Specifically, the preset adaptive path planning and delivery scheduling model adopts a multi-objective adaptive genetic algorithm architecture. The data flow consists of three layers: feature fusion and analysis layer, multi-objective population iterative optimization layer, and optimal solution selection and output layer. Multi-dimensional perturbation feature variables and baseline delivery path feature vectors are fused and analyzed layer by layer, multi-constraint iterative optimization is performed, and the optimal delivery solution is selected and output.

[0051] The pre-defined adaptive route planning and delivery scheduling model is built on an adaptive genetic algorithm. The feature fusion and analysis layer integrates perturbation variables and baseline path features and performs normalization processing. The optimization layer dynamically adjusts the population search width and the number of candidate solutions based on the perturbation level, and iteratively updates the population with delivery time, fresh produce loss, and path reliability as optimization objectives. The output layer compares all feasible solutions to select the comprehensive optimal route. During the training phase, the model adaptively calibrates by relying on historical delivery deviation data to correct crossover and mutation-related operating parameters.

[0052] The input to the pre-defined adaptive route planning and delivery scheduling model consists of two types of data: multi-dimensional perturbation feature variables and baseline delivery route feature vectors. The model outputs a complete final delivery route plan, which includes the access order of each stop, segmented driving routes, planned arrival time of each node, and planned stop duration. This plan is used to distribute the data to the vehicle terminal and conduct subsequent execution deviation collection and analysis.

[0053] In this embodiment, the final delivery route plan is a standardized scheduling file output after the preset adaptive route planning and delivery scheduling model is solved. It includes the complete vehicle driving route, the order of each stop, the planned arrival / stop time of each node, the temperature control requirements of the vehicle compartment, the suggested speed for each segment of driving, and the estimated fresh produce loss. It can be directly sent to the vehicle terminal and management monitoring platform for execution.

[0054] As can be seen from the above, this application delineates the spatiotemporal constraint region by considering orders, vehicles, and fresh produce deterioration characteristics. It constructs a disturbance feature vector using multi-source real-time data, quantifies the disturbance level of the delivery network through an evaluation model, and accurately locates local reconstruction of high-risk road segments, avoiding computational redundancy caused by global replanning. It integrates path deviation risk and multi-dimensional disturbance variables into an adaptive scheduling model, simultaneously considering delivery time, fresh produce preservation, and path reliability. It achieves dynamic local path optimization driven by disturbance hierarchy, accurately predicting path deviation and delay risks, and reducing fresh produce quality loss. Based on real-time road conditions and cargo status, it adaptively adjusts the optimization strategy, improving delivery network stability, shortening delivery delay time, balancing scheduling computation efficiency and fresh produce delivery quality, and enabling large-scale multi-vehicle cold chain real-time scheduling.

[0055] In one embodiment of this application, the current delivery route of the cold chain delivery vehicle is obtained, and based on the stability assessment results and disturbance level classification, target optimized route segments requiring route replanning or partial adjustment are determined from the current delivery route, including: Based on the preset perturbation type-weight mapping table, corresponding fuzzy evaluation weights are assigned to different identified perturbation types. The fuzzy comprehensive evaluation method is used to evaluate the various road segments along the current delivery route and generate a disturbance risk score curve along the route. The multiple factors include road congestion, temperature fluctuation, time window urgency, and vehicle remaining range. Based on the disturbance risk scoring curve, continuous road segments with scores greater than or equal to the preset risk scoring threshold are extracted to obtain high-risk road segments. Real-time traffic flow data and fresh produce quality decay rate corresponding to high-risk path segments are obtained. The start and end points of the high-risk path segments are corrected based on the traffic flow-decrease coupling model to obtain the target optimized path segment.

[0056] In this embodiment, the disturbance type-weight mapping table is a lookup table pre-stored on the scheduling platform. Each record corresponds to a type of delivery disturbance (road congestion, outdoor high temperature, abnormal vehicle temperature control, insufficient vehicle range, tight customer time window, etc.). The table pre-stores the fuzzy evaluation weight values ​​matched for each type of disturbance. When the disturbance level is different, the corresponding weight can be retrieved for fuzzy evaluation calculation. The fuzzy evaluation weight is the proportion coefficient assigned to each evaluation factor in the fuzzy comprehensive evaluation calculation, and the sum is 1. The greater the disturbance hazard, the greater the weight assigned, indicating that the indicator has a higher degree of impact on road segment risk.

[0057] In this embodiment, the fuzzy comprehensive evaluation method is a comprehensive scoring algorithm for handling multiple fuzzy indicators that are difficult to quantify precisely. First, the risk fuzziness level of each factor is divided, a membership function is constructed, the membership degree of each road segment indicator is calculated, and the comprehensive risk score is obtained by combining the weight matrix. This method is suitable for evaluation indicators with fuzzy boundaries such as congestion and temperature.

[0058] Specifically, the fuzzy comprehensive evaluation process consists of three data flow layers: the disturbance factor weight allocation layer, the single-factor membership calculation layer, and the multi-level comprehensive score output layer. The four types of evaluation indicators for each road segment complete weight matching, membership quantification, and multi-factor fusion calculation layer by layer to generate a continuous disturbance risk score curve for the entire delivery route.

[0059] First, a preset disturbance type-weight mapping table is retrieved to assign corresponding fuzzy evaluation weights to four factors: road congestion, temperature fluctuation, time window urgency, and remaining vehicle range. Then, membership functions for each indicator are constructed for a single road segment, converting the actual values ​​of the indicators into membership degrees between 0 and 1. The weight matrix and membership matrix are combined using fuzzy operators to output the comprehensive risk score segment by segment. The scores of all road segments are then concatenated along the driving path to form a continuous disturbance risk scoring curve.

[0060] The input consists of four measured indicators for each segment of the current delivery route: traffic congestion, temperature fluctuation, time window urgency, and remaining vehicle range, along with a disturbance type-weight mapping table. The output is a disturbance risk score curve covering the entire delivery route, used to filter out continuous high-risk route segments whose scores exceed the preset risk score threshold, providing a basis for subsequent corrections to obtain the target optimized route segment.

[0061] In this embodiment, road congestion level is a quantitative indicator of real-time traffic congestion on a road segment, ranging from 0 to 1. A higher value indicates slower vehicle movement and longer waiting times, directly extending the exposure time of fresh produce at room temperature. Temperature fluctuation amplitude is the short-term difference between outdoor ambient temperature and vehicle refrigeration temperature; greater fluctuation indicates more intense heat exchange in the cold chain, significantly accelerating the deterioration of fresh produce quality, making it a unique evaluation indicator for the cold chain. Time window urgency is a quantitative value of the remaining buffer time for the customer's requested delivery period; a higher value indicates closer proximity to the latest delivery time, making route delays highly likely to result in late delivery and breach of contract. Vehicle remaining range is the maximum driving distance that the vehicle's remaining battery / fuel capacity can support. If the road segment is too long or subject to significant disturbances, insufficient range may cause the vehicle to stop midway and heat up, exacerbating the spoilage of fresh produce.

[0062] In this embodiment, the disturbance risk scoring curve is a continuous curve plotted with the delivery route's travel sequence on the horizontal axis and the comprehensive risk score of each segment on the vertical axis, visually displaying the risk distribution along the entire route. The preset risk scoring threshold is a calibrated critical risk score; if the comprehensive score of a segment is greater than or equal to this value, it is determined to be high-risk, requiring route adjustments; if it is less than the preset risk scoring threshold, no changes are needed.

[0063] The preset risk score threshold is determined using a historical delivery sample statistical calibration method. Specifically, three months of complete cold chain delivery historical data for the target delivery area are collected first. Four evaluation indicators are extracted for each road segment: road congestion, temperature fluctuation range, time window urgency, and vehicle remaining range. The disturbance risk score for each historical road segment is calculated through fuzzy comprehensive evaluation. The fresh produce loss rate and delivery timeout probability corresponding to different scores are statistically analyzed, and a risk score-loss timeout correlation curve is plotted. Using the maximum acceptable fresh produce loss threshold and the order timeout tolerance ratio as constraints, the corresponding critical risk score is calculated inversely from the correlation curve as the initial threshold. Then, 10 sets of test samples under extreme congestion and high temperature disturbance scenarios are selected for verification, and the initial critical score is corrected to finally obtain the preset risk score threshold.

[0064] In this embodiment, the maximum fresh produce loss threshold is jointly calibrated using a combination of physicochemical testing of fresh produce categories and operational cost loss calculation: Constant temperature, variable temperature, and delayed transportation comparative tests are conducted on various types of fresh produce covered by the delivery service to determine the appearance, taste, and shelf-readiness standards corresponding to different levels of loss, thus identifying the critical loss coefficient corresponding to spoilage and unsaleable fresh produce. Combined with the company's fresh produce procurement costs, after-sales compensation costs, and customer churn losses, the comprehensive economic loss corresponding to different loss rates is calculated. Using the company's preset maximum acceptable comprehensive loss limit for a single batch of delivery as a constraint, the critical loss coefficient obtained from the tests is matched to obtain the initial maximum fresh produce loss threshold. This is then corrected by combining delivery samples from different seasons, such as summer high temperatures and winter low temperatures in the target area, ultimately yielding the maximum fresh produce loss threshold by category and season. The order overtime tolerance ratio is the percentage of orders that are allowed to be delivered late; exceeding this ratio will result in customer complaints and compensation losses.

[0065] In this embodiment, the high-risk path segment is a continuous adjacent road section on the current delivery route that all meet the risk score greater than or equal to the threshold. It has a uniform high-disturbance characteristic and is suitable for unified local reconstruction.

[0066] In this embodiment, real-time traffic flow data represents the number of vehicles passing through high-risk road sections per unit time. Higher traffic flow indicates more severe congestion and longer time for fresh produce to remain on the road. The fresh produce quality degradation rate is a quantification coefficient of the quality loss of a corresponding category of fresh produce per unit travel time under the current ambient temperature; higher temperatures and longer periods of congestion result in a higher degradation rate.

[0067] In this embodiment, the flow-attenuation coupling model is a constructed correlation calculation model. It takes the traffic flow and fresh produce attenuation rate as inputs and outputs the additional freshness loss caused by congestion. Based on the loss increment, it corrects the start and end nodes of high-risk road sections, eliminates small road segments with weak loss impact, and accurately locates the core high-loss interval.

[0068] Specifically, the flow-attenuation coupling model data flows through three levels: the data standardization input layer, the flow and quality attenuation coupling calculation layer, and the road segment start and end point correction output layer. Real-time traffic flow and fresh produce quality attenuation rate are processed layer by layer to achieve dimensional unification, coupling correlation calculation, and road segment interval correction, outputting a precise target optimized path segment interval.

[0069] The flow-deterioration coupling model first eliminates the dimensional differences between traffic flow and quality deterioration rate through a standardization layer; the coupling operation layer establishes a correlation function between traffic congestion duration and fresh produce spoilage rate, quantifying the cumulative impact of traffic congestion delays on the spoilage rate of cold chain goods; based on the comprehensive loss risk coefficient obtained from the coupling calculation, the original start and end nodes of high-risk path segments are shifted, contracted, or extended to complete the interval correction; the entire model calibrates the coupling coefficient based on a large number of historical samples of road traffic flow and cold chain loss.

[0070] The flow-attenuation coupling model takes real-time traffic flow data and fresh goods quality attenuation rate as inputs for high-risk path segments; the output is the target optimized path segment after origin-end point correction, which is used for subsequent local path reconstruction and adjustment to generate the reconstructed path segment.

[0071] In this embodiment, the start and end points of the high-risk path segment are the adjacent delivery stop nodes at the beginning and end of the high-risk path segment. Initially, these are determined solely based on risk scores, without any adjustments for fresh produce spoilage. The target optimized path segment is the final high-risk continuous path segment after the start and end nodes have been corrected using a flow-attenuation coupling model; it is the object of local reconstruction.

[0072] As can be seen from the above, this embodiment, by introducing disturbance type weight mapping and fuzzy comprehensive evaluation method, integrates multi-dimensional indicators such as congestion, temperature, time window, and range to generate a continuous risk score curve, which can accurately identify high-risk road sections along the entire route. By correcting the start and end nodes of road sections through a traffic flow and fresh produce deterioration coupling model, the target route segments that need optimization are accurately identified. It achieves multi-factor quantitative fuzzy classification, risk identification covers the unique preservation constraints of cold chain delivery, distinguishes between high and low risk road sections to avoid indiscriminate route adjustments; the traffic flow-deterioration coupling correction logic links traffic congestion and cargo loss in calculation, improving the accuracy of target optimization road section selection, reducing invalid route reconstruction operations, lowering scheduling computing power consumption, and suppressing the quality deterioration of fresh produce caused by high-risk road sections.

[0073] In one embodiment of this application, a fuzzy comprehensive evaluation method is used to perform multi-factor assessments on each segment along the current delivery route, generating a disturbance risk score curve along the route, including: The current delivery route is divided into several continuous segments according to the interval between adjacent stops, and the measured factor values ​​of each segment are obtained. Each evaluation factor for each road segment is divided into several fuzzy levels, including a first risk level, a second risk level, and a third risk level. The first risk level is lower than the second risk level, and the second risk level is lower than the third risk level. Construct membership functions for each factor to each fuzzy level, using either a trapezoidal or triangular distribution for the membership functions; The membership matrix corresponding to each road segment is calculated based on the measured factor values ​​of each road segment. The membership matrix is ​​then combined with the fuzzy evaluation weights to obtain the membership vector of each road segment to each risk level. Extract the membership value of the third risk level from the membership vector as the disturbance risk score of the corresponding road segment, and connect the disturbance risk scores of all road segments in the order of the paths to generate a disturbance risk score curve.

[0074] In this embodiment, the interval between adjacent stops is the complete road section between two consecutive receiving / unloading points during the vehicle's delivery route. This serves as the standard for road segmentation, aligning with the business logic of multi-point cold chain delivery. Continuous road segments are independent, non-overlapping, and sequentially connected road units after segmentation, with each segment undergoing independent risk assessment. The measured factor values ​​are the raw values ​​of four evaluation indicators collected in real time: road congestion, temperature fluctuation range, time window urgency, and remaining vehicle range.

[0075] In this embodiment, the evaluation factors are four dimensions used to determine the risk of road segment disturbance: road congestion, temperature fluctuation range, time window urgency, and vehicle remaining range. The fuzzy rating is a three-tiered qualitative classification of the risk level of a single evaluation factor, without clear hard boundaries, relying on membership measurement to quantify the degree of affiliation. The first risk level is low risk, with slight disturbance and almost no impact on fresh produce and timeliness; the second risk level is medium risk, with some potential for loss and delay; the third risk level is high risk, which can easily cause fresh produce spoilage and order delays.

[0076] In this embodiment, the membership function is a core calculation function of fuzzy mathematics. It takes a measured value of the indicator as input and outputs a membership value between 0 and 1, representing the degree to which the value belongs to a certain risk level. It includes two commonly used forms: triangular distribution and trapezoidal distribution. The triangular distribution membership function has a triangular graph and reaches a membership degree of 1 only in a single interval. It is suitable for indicators with clear boundaries, such as the remaining range of a vehicle. The trapezoidal distribution membership function has a trapezoidal graph and a membership degree that is stable at 1 within a certain interval. It is suitable for indicators with interval fluctuations, such as ambient temperature fluctuations and road congestion.

[0077] In this embodiment, the membership matrix is ​​a 4x3 matrix composed of the membership values ​​of the four evaluation factors corresponding to the three fuzzy levels under a single road segment. The fuzzy evaluation weights are the weight coefficients assigned to the four evaluation factors, summing to 1, representing the proportion of each indicator's influence on the overall risk. The synthesis operation is a fuzzy matrix weighted operation, multiplying the membership matrix by the weight vector to achieve a multi-indicator fusion and comprehensive scoring. The membership vector is a one-dimensional array obtained after the synthesis operation, containing three values, representing the membership degree of the road segment belonging to the first, second, and third risk levels, respectively.

[0078] In this embodiment, the disturbance risk score is directly taken as the membership value corresponding to the third risk level (high risk), with a value of 0-1. The higher the value, the greater the risk of the road segment. The disturbance risk score curve is a continuous numerical sequence formed by the delivery route driving sequence as the horizontal axis and the disturbance risk score of each road segment as the vertical axis, representing the risk change trend throughout the entire process.

[0079] As can be seen from the above, this embodiment divides the delivery route into independent segments according to the stop points, constructs a multi-factor fuzzy evaluation system using trapezoidal / triangular membership functions, divides the risk into three levels, and quantifies the disturbance risk of each segment by weighted synthesis of the membership matrix, generating a continuous risk curve. This solves the shortcomings of existing risk assessments that cannot quantify multi-dimensional fuzzy indicators, adapting to evaluation factors such as road conditions, temperature, and time windows that cannot be precisely numerically defined; the segmented evaluation aligns with cold chain multi-point delivery operations, generating a visual risk curve according to the route sequence, which can intuitively represent the risk distribution of the entire route; the unified membership calculation standard ensures the consistency of risk scoring, providing stable and quantifiable data support for high-risk segment selection, and improving the objectivity and accuracy of target optimization segment extraction.

[0080] In one embodiment of this application, the target optimized path segment is partially reconstructed and adjusted to generate a reconstructed path segment. The reconstructed path segment is then concatenated with the unadjusted portion of the current delivery path, and features are extracted to generate a second delivery path feature vector, including: Obtain the start and end node information of the target optimized path segment and the corresponding constraints. The constraints include the stop time window, vehicle remaining capacity, remaining shelf life of fresh goods, and collected road condition data and ambient temperature data. With the optimization objectives of minimizing delivery time, minimizing cargo quality degradation, and maximizing path reliability, a local optimization algorithm is used to replan feasible paths between start and end nodes, generating reconstructed path segments. The reconstructed route segment is then combined with the unadjusted portion of the current delivery route to form the updated complete delivery route. Feature extraction is performed on the updated complete delivery route to obtain the feature vector of the second delivery route.

[0081] In this embodiment, the target optimized path segment is a continuous high-risk road segment after the start and end nodes have been corrected by the flow-attenuation coupling model. It is the section where the road needs to be replanned. The start and end node information consists of the delivery stop number, geographical location, and delivery order information of the two adjacent delivery stop points at both ends of the target optimized path segment. These remain fixed during reconstruction, and only the road between the two points is changed.

[0082] In this embodiment, the constraints are the hard restrictions on cold chain delivery that must be followed during local route reconstruction. If any constraint is not met, the generated route is determined to be infeasible. Among them, the stop time window is the start and end time period for receiving goods specified by the customer, and the route arrival time cannot exceed this interval; the vehicle remaining capacity is the remaining cargo space / load limit of the carriage, and the reconstructed route cannot add cargo midway; the remaining freshness time of fresh goods is the longest transportation time that the fresh goods in the current carriage can withstand, and exceeding this time will result in a large number of spoilages and scrapping; the road condition data is the real-time congestion, construction, and speed limit information within the interval; the ambient temperature data is the real-time outdoor temperature of the road section, and high-temperature road sections are given priority to be avoided.

[0083] In this embodiment, the optimization objectives are three evaluation criteria followed by the local optimization algorithm during iterative solution. Minimizing delivery time is to shorten the travel time between the start and end nodes and reduce the exposure time of fresh produce. Minimizing the degradation of goods quality is to prioritize low-temperature and smooth road sections to reduce the rate of spoilage of fresh produce. Maximizing path reliability is to avoid unstable roads with temporary construction, high congestion, or high temperature exposure.

[0084] In this embodiment, the local optimization algorithm is an optimization algorithm that searches for feasible routes only within the road network between fixed start and end nodes. This embodiment uses the particle swarm optimization algorithm without changing the stop order of the entire route. A feasible path is a route that simultaneously satisfies all time windows, vehicle capacity, fresh produce preservation, and temperature and road condition constraints, and has no violations, timeouts, or excessive losses.

[0085] Specifically, the local optimization algorithm in this embodiment adopts the particle swarm optimization algorithm, which has three layers of data flow: the solution space and hyperparameter configuration layer, the particle iterative optimization operation layer, and the optimal feasible path output layer. The start and end node constraints, road network topology, and disturbance feature data are used to complete parameter initialization, multi-objective iterative update, and optimal path selection layer by layer, and output the reconstructed path segment.

[0086] The configuration layer defines the particle solution space based on the road network topology and the number of stops. The particle position vector corresponds to the access order of intermediate stops. Then, it combines the disturbance level, road condition data to match hyperparameters such as inertia weight and acceleration factor. The optimization layer constructs a fitness function based on delivery time, freshness decay, and route reliability. It continuously updates the particle speed and position and verifies the constraints of time window, cargo capacity, and freshness duration. After iterative convergence, the route corresponding to the particle with the best fitness is selected as the reconstructed path segment.

[0087] The input to the particle swarm optimization algorithm includes the start and end nodes of the target optimized path segment, multiple constraints such as time window / vehicle capacity / shelf life / road condition temperature, road network topology, disturbance level classification, and road condition data with the first disturbance feature vector. The output is the reconstructed path segment that satisfies all constraints between the start and end nodes, which is used to splice with the unadjusted part of the original path and generate the second delivery path feature vector after feature extraction.

[0088] In this embodiment, the reconstructed path segment is a completely new route obtained through algorithmic optimization, replacing the original high-risk road segments, with the start and end nodes remaining consistent with the original road segment. The unadjusted portion of the current delivery route is all road segments, stop order, and driving plan of the original delivery route excluding the target optimized route segment; these are completely preserved without modification during splicing. Splicing involves replacing the target optimized interval in the original path with the reconstructed new road segment, connecting the two unadjusted route segments to form a continuous, complete route with the stop order unchanged.

[0089] In this embodiment, the updated complete delivery route is a complete delivery plan with all delivery nodes and driving roads updated after replacing high-risk road sections. Feature extraction involves extracting quantitative values ​​such as mileage, total delivery time, temperature control fluctuations, and time window tightness. The second delivery route feature vector is a one-dimensional standardized numerical array generated by arranging all features of the updated complete route in a fixed order. Its dimension and sorting rules are completely consistent with the baseline delivery route feature vector, and it is used for path deviation identification and comparison.

[0090] As can be seen from the above, this embodiment incorporates multiple constraints such as time window, vehicle capacity, freshness preservation time, and temperature and road conditions during route reconstruction. It collaboratively optimizes multiple objectives, including delivery time, freshness decay, and route reliability. After splicing the old and new routes, it extracts the complete route feature vector for deviation identification. Unlike single-objective route reconstruction that only shortens the driving distance, this embodiment also takes into account the cold chain preservation requirements. The constraints include all hard business rules for fresh food delivery, preventing issues such as timeouts, overloads, and freshness failures in the reconstructed route. The partial reconstruction splicing mode does not require recalculating all delivery nodes, reducing the computational workload of route adjustment. The generation of a second route feature vector enables quantitative comparison of the old and new routes, accurately identifies the risk of new deviations after reconstruction, ensures the overall stability of the optimized route, and effectively balances scheduling efficiency and fresh food delivery quality.

[0091] In one embodiment of this application, the local optimization algorithm employs particle swarm optimization; the local optimization algorithm is used to replan feasible paths between the start and end nodes, generating reconstructed path segments, including: Based on the road network topology and the number of stops between the start and end nodes of the target optimized path segment, the solution space of the particle swarm optimization algorithm is determined. The position vector of each particle in the solution space represents a visiting order of intermediate stops between the start and end nodes. Based on the disturbance level classification and the road condition data in the first disturbance feature vector, the hyperparameter combination of the particle swarm optimization algorithm is determined.

[0092] In this embodiment, the local optimization algorithm searches the route only within the fixed first and last nodes of the target optimized path segment, without changing the order of other stops along the entire delivery route. This embodiment employs the particle swarm optimization algorithm. The particle swarm optimization algorithm is an intelligent optimization algorithm based on population iterative optimization. It searches for the optimal solution by moving a swarm of particles within the solution space and is suitable for combinatorial optimization problems involving path node ordering.

[0093] In this embodiment, the road network topology between the start and end nodes is the connection relationship of all main and secondary roads, branch roads, and interchanges within the scope of the start and end nodes of the target optimized road segment, including road connectivity, intersection turning constraints, and road traffic restrictions, which constitute the basic boundary of the optimized road network.

[0094] In this embodiment, the number of stops is the total number of intermediate pickup points such as communities, shops, and self-pickup points distributed between the start and end nodes, which determines the dimensionality of the path sorting solution. The solution space is the set of all legal and feasible solutions of the particle swarm optimization algorithm. It is constrained by the road network connectivity rules and the number of intermediate stops, and only includes node arrangement schemes that allow normal passage between two points without reverse traffic restrictions, eliminating invalid routes that cannot be passed.

[0095] In this embodiment, particles are the basic search units in the particle swarm optimization algorithm, with each particle representing a set of alternative intermediate stop access schemes. The particle's position vector is a one-dimensional array, with the array length equal to the number of intermediate stops between the start and end nodes. The elements in the array are stop numbers, and the array order corresponds to the order in which vehicles access intermediate stops. Disturbance levels are classified into minor disturbances and severe disturbances, representing the overall strength of delivery interference. The first disturbance feature vector incorporates quantitative road condition indicators such as road segment congestion index, travel delay, and outdoor temperature. Road condition data are the feature components in the first disturbance feature vector that characterize traffic flow status, including quantitative values ​​such as congestion degree, average travel delay, and peak traffic flow percentage.

[0096] In this embodiment, the hyperparameter combination is a complete set of control parameters that need to be preset before the particle swarm algorithm runs, including inertia weight, individual learning factor, global learning factor, population size, maximum number of iterations, and upper limit of particle search speed; different parameter combinations are matched for different perturbation scenarios.

[0097] As can be seen from the above, this embodiment defines the particle swarm optimization solution space based on road topology and the number of stops, with particle position vectors directly corresponding to the intermediate point access order, which aligns with the needs of cold chain multi-point delivery route planning. The particle swarm hyperparameter combination is dynamically configured based on disturbance level and real-time road conditions. Existing particle swarm algorithms use fixed parameters, which are prone to getting stuck in local optima under severe disturbances and have low search efficiency under slight disturbances. This embodiment associates the road network size with the disturbance state to determine the solution space range, adapting to the optimization needs of road segments of different lengths. This achieves deep adaptation of the particle swarm algorithm to the dynamic disturbance scenario of the cold chain, improving the convergence speed of local path reconstruction optimization. Under severe disturbances, it broadens the search range to avoid high-risk road segments, and under slight disturbances, it quickly converges to the optimal solution, balancing optimization accuracy and computational speed.

[0098] In one embodiment of this application, the hyperparameters include inertial weights and acceleration factors; Based on the disturbance level classification and the road condition data in the first disturbance feature vector, the hyperparameter combination of the particle swarm optimization algorithm is determined, including: Obtain the disturbance level classification, which includes severe disturbances and minor disturbances; If the disturbance level is classified as a severe disturbance, the reference value of the inertia weight is increased based on the first preset step size, and the reference value of the acceleration factor is decreased based on the second preset step size. If the disturbance level is classified as a minor disturbance, the reference value of the inertia weight is reduced based on the third preset step size, and the reference value of the acceleration factor is increased based on the fourth preset step size.

[0099] In this embodiment, hyperparameters are control parameters configured before the particle swarm optimization algorithm runs and not automatically updated during the iteration process. In this embodiment, there are two types of parameters: inertia weight and acceleration factor. Inertia weight is a coefficient that controls how well particles retain the search speed and historical search direction from the previous round. A larger value indicates a stronger global exploration capability, suitable for severely disturbed scenarios with large-scale congestion and many high-risk branches; a smaller value indicates a faster convergence speed, suitable for slightly disturbed scenarios with stable road conditions. Acceleration factor includes individual learning factor and global learning factor, collectively called acceleration factor. A higher value indicates that particles quickly approach the current optimal route and have a stronger ability for local fine-tuning; a lower value indicates that particles traverse more of the global candidate paths.

[0100] In this embodiment, the disturbance level classification is a qualitative classification output by the delivery status assessment model, which is divided into severe disturbance and slight disturbance, and serves as a decision switch for parameter adjustment.

[0101] In this embodiment, the first preset step size is the increment of a single upward adjustment of the inertial weight under severe perturbation, and is a preset constant used to uniformly increase the global search capability. The second preset step size is the decrement of a single downward adjustment of the acceleration factor under severe perturbation, and is a preset constant.

[0102] In this embodiment, the third preset step size is the decrease in inertial weight under slight perturbation. The fourth preset step size is the increment in acceleration factor under slight perturbation.

[0103] In this embodiment, the reference values ​​are the pre-calibrated baseline values ​​of inertia weight and acceleration factor under the standard operating condition without disturbance, and all increment / decrement steps are calculated based on these baseline values ​​to obtain the final parameters.

[0104] The first preset step size is the incremental increase of the inertial weight under severe disturbance scenarios. It is jointly calibrated by multiple sets of disturbance comparison simulation experiments combined with real-time scheduling delay constraints: multiple sets of cold chain delivery simulation road networks covering mild congestion, large-area road congestion, and high-temperature multi-segment disturbances are used; the acceleration factor, population size, and number of iterations are fixed, and other parameters remain unchanged, only the gradient is set with different inertial weights to increase the step size, and local particle swarm optimization is run in batches; three indicators corresponding to each step size are recorded: global route search completeness, algorithm single optimization operation time, and accuracy of optimal path identification for high-risk road sections; with the dual constraints of not missing low-loss smooth routes under severe disturbances and the single optimization delay not exceeding 150ms of the vehicle scheduling system, the smallest increment that simultaneously satisfies search sufficiency and real-time is selected as the initial first preset step size; then, two types of real extreme disturbance samples, morning peak full-area congestion and summer high-temperature construction, are selected for retesting, and the values ​​are fine-tuned to obtain the first preset step size.

[0105] The second preset step size is the reduction of the acceleration factor under severe disturbance scenarios. It is jointly calibrated by multi-condition simulation control test and path optimization convergence index: through a simulated road network with severe disturbances including main road congestion and construction on road sections exposed to high temperatures, the inertia weight, population size, number of iterations and other parameters are fixed, and multiple sets of acceleration factor reduction step sizes are set to carry out batch local particle swarm optimization. The global feasible path traversal coverage, algorithm convergence iteration steps and optimal low-loss path detection rate corresponding to each set of step sizes are collected. The initial reduction step size is selected with the constraint that it can fully traverse detour branches under severe disturbances and will not quickly converge to a local suboptimal route, while meeting the real-time computing latency requirements of the vehicle terminal. Then, it is retested and verified using real extreme delivery samples such as morning peak full-area congestion and continuous high temperature in summer. After fine-tuning the values, the second preset step size is obtained.

[0106] The third preset step size is the reduction of inertial weights under slight disturbance scenarios. It is jointly calibrated by a stable road condition simulation control test combined with vehicle computing power delay constraints: through a simulated road network with slight disturbances such as smooth traffic flow, no high temperature exposure, and no road construction, the acceleration factor, population size, maximum number of iterations, and other particle swarm parameters are kept unchanged, and multiple sets of inertial weight reduction step sizes are set to carry out batch local path optimization; the number of algorithm convergence iterations, single optimization calculation time, and optimal path matching accuracy corresponding to each set of step sizes are collected; with the constraints of fast convergence under slight disturbance conditions, reduced vehicle computing power consumption, and no loss of the best short-distance smooth route, the initial reduction step size is selected; then, multiple sets of real low-disturbance delivery samples such as off-peak commuting and spring and autumn normal temperature are retrieved for retesting and verification, and the values ​​are corrected to obtain the third preset step size.

[0107] The fourth preset step size is the incremental increase of the acceleration factor under slight disturbance scenarios. It is jointly calibrated by a stable road condition simulation comparison test and the path convergence efficiency index. A slight disturbance simulation road network with smooth traffic flow, suitable temperature, and no road construction control is constructed. The inertial weight benchmark value, population size, maximum number of iterations, and other particle swarm parameters are kept unchanged. Multiple sets of acceleration factor increase step sizes are set to batch execute local path optimization. The number of algorithm convergence iterations, the time consumption of a single optimization operation, and the accuracy of the identification of the best short and smooth path are collected for each set of step sizes. The initial increase step size is selected with the dual constraints of accelerating particles to move closer to the local optimal route under slight disturbance, shortening the iteration time, and not getting trapped in local inferior routes. Then, multiple sets of real slight disturbance samples such as spring and autumn off-peak suburban delivery and normal temperature low congestion are retrieved for retesting and verification. After correcting the values, the fourth preset step size is obtained.

[0108] As can be seen from the above, this embodiment differentiates the inertia weight and acceleration factor by distinguishing between severe and minor disturbance levels. Severe disturbances increase the inertia weight to enhance global search capabilities and decrease the acceleration factor to reduce local extreme convergence; minor disturbances decrease the inertia weight and increase the acceleration factor to accelerate local optimal convergence. Fixed hyperparameter particle swarm optimization is prone to convergence to suboptimal paths in scenarios with complex road conditions and severe disturbances, exacerbating fresh produce spoilage. This embodiment achieves adaptive dynamic adjustment of algorithm hyperparameters according to the disturbance intensity. When the disturbance is severe, it fully traverses feasible routes to avoid temperature and congestion risks, and when the disturbance is mild, it quickly locks the optimal short-distance route, improving the overall optimization effect of local path reconstruction, reducing delivery delays and fresh produce quality degradation, and shortening the optimization iteration time.

[0109] In one embodiment of this application, after inputting multi-dimensional perturbation feature variables and baseline delivery path feature vectors into a preset adaptive path planning and delivery scheduling model to obtain the optimized final delivery path scheme, the method further includes: The final delivery route plan is sent to the on-board terminal of the corresponding cold chain delivery vehicle and pushed to the monitoring dashboard of the delivery management platform. Collect actual operational data of cold chain delivery vehicles during the delivery process according to the final delivery route plan. The actual operational data includes the actual driving trajectory, actual stopping time and actual arrival time of each node. The actual operational data is compared with the expected data in the final delivery route plan to generate a plan execution deviation report; Based on the deviation report of the plan execution, adjust the algorithm running parameters in the adaptive route planning and delivery scheduling model.

[0110] In this embodiment, the final delivery route plan is the complete delivery execution plan, including the complete driving route, the order of visits to each stop, the expected arrival time of each node, the expected stop duration, the expected vehicle speed for each segment, the estimated spoilage of fresh produce, and other complete expected scheduling data. The vehicle terminal is an intelligent scheduling device installed on the cold chain delivery vehicle, which has the functions of route navigation, task reception, and real-time data upload. It is used to receive and display the issued delivery route and record data of the entire vehicle driving process.

[0111] In this embodiment, the delivery management platform is a cloud-based backend dispatch system that centrally manages all cold chain vehicle delivery tasks. The monitoring dashboard is the platform's visual interface, allowing dispatchers to view all vehicle planned routes, real-time locations, and delivery progress in real time. The monitoring dashboard is the visual interface of the delivery management platform, visually displaying vehicle locations, planned routes, time-sensitive alerts, and estimated fresh produce spoilage, facilitating real-time monitoring of the entire delivery process by dispatchers.

[0112] In this embodiment, the actual operating data is the real-time working condition data transmitted back by the vehicle terminal during the delivery task, including three types of core information: actual driving trajectory: the actual driving road formed by the continuous GPS positioning points of the vehicle; actual stopping time: the actual time the vehicle stops at each receiving node for loading and unloading; and actual arrival time of each node: the actual time the vehicle arrives at each delivery stop.

[0113] In this embodiment, the expected data refers to the theoretically planned values ​​stored in the final delivery route plan, including the expected driving trajectory, the expected stop time at each node, the expected arrival time, and the segmented travel time. The plan execution deviation report is a quantitative document generated by calculating the differences between the actual operating data and the expected data item by item. It includes quantitative indicators such as route spatial offset, arrival time deviation, stop time deviation, total timeliness deviation, and timeout frequency at each node, which intuitively represent the error in the implementation of the route plan.

[0114] In this embodiment, the adaptive path planning and delivery scheduling model incorporates adjustable algorithm parameters such as population size, adaptive crossover probability, and mutation probability. These algorithm parameters are adjustable control parameters of the adaptive path planning and delivery scheduling model, dynamically fine-tuned based on the magnitude of execution deviation to reduce the error between subsequent planning and actual delivery.

[0115] As can be seen from the above, this embodiment achieves two-way visual control by simultaneously distributing optimized routes to vehicle terminals and management dashboards. It collects actual operational data throughout the entire process, compares it with the planned scheme to generate deviation reports, and iterates the scheduling model parameters through a closed-loop process using deviation data. Existing cold chain optimization schemes only output planned routes and lack feedback during execution and model self-correction mechanisms, leading to the continuous accumulation of planning deviations over time. This embodiment constructs a complete closed loop of route distribution, execution data collection, deviation analysis, and parameter correction. It monitors vehicle trajectories, stops, and arrival time deviations in real time, adjusts the model search strategy based on actual delivery execution, continuously improves the accuracy of subsequent route planning, reduces delivery delays and trajectory deviations, and minimizes fresh produce spoilage and scheduling operation costs.

[0116] In one embodiment of this application, the actual operating data is compared with the expected data in the final delivery route plan to generate a plan execution deviation report, including: Obtain the actual driving trajectory, actual stopping time, and actual arrival time of each node from the actual operation data; The actual driving trajectory is spatially aligned with the planned route in the final delivery route plan. The deviation distance between the actual driving route and the planned route is calculated for each road segment. Road segments with a deviation distance greater than or equal to a preset deviation threshold are marked as trajectory deviation road segments. The actual stopping time is compared with the planned stopping time in the final delivery route plan, and the stopping time deviation value of each stopping point is calculated. The stopping time deviation value is the actual stopping time minus the planned stopping time. The actual arrival time of each node is compared with the planned arrival time in the final delivery route plan, and the arrival time deviation value of each node is calculated. The arrival time deviation value is the actual arrival time minus the planned arrival time. Based on the deviations in trajectory, stopping time, and arrival time, a plan execution deviation report is generated, including the deviation type, deviation location, deviation amount, and deviation duration.

[0117] In this embodiment, the actual driving trajectory is a sequence of location coordinates continuously collected by the vehicle's GPS, which is then strung together to form the complete road route the vehicle actually travels. The actual stopping time is the total duration for which the vehicle actually stops at each delivery node to load and unload goods. The actual arrival time at each node is the precise time when the vehicle actually arrives at each receiving stop.

[0118] In this embodiment, the planned path is a pre-set standard driving route within the final delivery route plan. Spatial alignment unifies the actual trajectory coordinates and the planned path coordinates into the same geographic coordinate system, completing one-to-one matching of road segments and enabling deviation comparison within the same road segment. Deviation distance is the average vertical distance between the vehicle's actual driving position and the centerline of the planned route within the same road segment, characterizing the degree of route deviation. The preset deviation threshold is a calibrated distance critical value; if the deviation distance is greater than or equal to this value, the vehicle is determined to have significantly deviated from the planned route, and this is recorded as a trajectory deviation segment. A trajectory deviation segment is a road section where the deviation distance is greater than or equal to the preset deviation threshold, representing abnormal driving behaviors such as the vehicle actively detouring or taking the wrong route.

[0119] The preset deviation threshold is determined by combining road network width surveying with the tolerance range for normal detours: Road width and lane distribution data of all main and secondary roads and community side roads in the delivery area are collected, and the vertical distance intervals of compliant small deviations such as normal obstacle avoidance and temporary roadside stops are statistically analyzed. Two scenarios are distinguished: one is normal small deviations (temporary avoidance of non-motorized vehicles, brief roadside stops), and the other is large deviations such as active detours, wrong road trips, and unauthorized route changes. Historical GPS trajectories of cold chain delivery vehicles over the past three months are statistically analyzed, and the vertical deviation distance between the trajectory and the planned route is calculated in batches to plot the frequency distribution curve of the deviation distance. With the constraint of filtering only abnormal active detours and not identifying temporary avoidance as abnormal deviations, the upper limit of the compliant deviation interval in the frequency curve is selected as the initial deviation threshold. Then, samples of complex road conditions such as narrow community side roads and multi-intersections are selected for retesting and verification, and the values ​​are corrected to obtain the preset deviation threshold.

[0120] In this embodiment, the planned stop time is the standard loading and unloading dwell time at each node pre-calculated in the final delivery route plan. The formula for calculating the stop time deviation is: actual stop time - planned stop time; a positive value indicates unloading timeout, and a negative value indicates early completion of loading and unloading. The planned arrival time is the preset standard time for the vehicle to arrive at each delivery node. The formula for calculating the arrival time deviation is: actual arrival time - planned arrival time; a positive number indicates delivery delay, and a negative number indicates early arrival.

[0121] In this embodiment, deviation types are divided into three categories: trajectory deviation, stop duration deviation, and node arrival time deviation. The deviation location is the road segment number, stop name, and geographical range corresponding to the deviation occurrence. The deviation quantity is a quantified value, including trajectory offset distance, stop time difference, and arrival time difference.

[0122] In this embodiment, the duration of deviation is the travel time of the deviated section of the route, or the cumulative duration of the delay caused by the deviation in stopping or arrival. The plan execution deviation report is a standardized document that integrates all quantitative deviation information and fully records all types of execution errors in the entire delivery process.

[0123] As can be seen from the above, this embodiment quantifies execution deviations from three dimensions: spatial trajectory, stop duration, and node arrival time. It calculates trajectory deviation distance through spatial alignment, calculates stop and arrival time deviations separately, and integrates multiple types of deviations to generate a standardized execution report. Existing deviation statistics only count arrival delay time, failing to distinguish between different risk factors such as route deviation and pickup timeouts. This embodiment achieves full-dimensional quantitative labeling of delivery deviations, accurately locating deviation segments, deviation types, and durations, intuitively representing the root causes of differences between route planning and actual execution, and providing refined, multi-dimensional quantitative data support for model parameter adjustment.

[0124] In one embodiment of this application, adjusting the algorithm operating parameters in the adaptive route planning and delivery scheduling model based on the scheme execution deviation report includes: Extract the cumulative length of the trajectory deviation segment, the mean and variance of the stop time deviation, and the mean and variance of the arrival time deviation from the plan execution deviation report; The path execution deviation rate is calculated based on the ratio of the cumulative length of the deviation segments to the total length of the planned path in the final delivery route plan. Calculate the docking reliability score based on the mean and variance of the docking time deviation. Calculate the timeliness reliability score based on the mean and variance of the arrival time deviations; The path execution deviation rate, dock reliability score and timeliness reliability score are weighted and fused to obtain the model confidence adjustment factor; The model confidence adjustment factor is smoothed using a moving average filter to obtain the smoothing adjustment amount. The smoothing adjustment amount is compared with a preset adjustment threshold, which includes a first adjustment threshold and a second adjustment threshold, wherein the first adjustment threshold is greater than the second adjustment threshold. If the smoothing adjustment amount is greater than or equal to the first adjustment threshold, the path search width and the number of candidate solutions in the adaptive path planning and delivery scheduling model are increased based on the fifth preset step size. If the smoothing adjustment amount is less than the second adjustment threshold, the path search width and the number of candidate solutions in the adaptive path planning and delivery scheduling model are reduced based on the sixth preset step size. If the smoothing adjustment is greater than or equal to the second adjustment threshold and less than the first adjustment threshold, then the current algorithm running parameters remain unchanged.

[0125] In this embodiment, the cumulative length of the trajectory deviation segment is the sum of the mileage of all segments in the scheme execution deviation report whose offset distance is greater than or equal to the preset deviation threshold, representing the degree to which the vehicle deviates from the planned route as a whole.

[0126] In this embodiment, the mean of the docking time deviation is the average level of the docking time deviation of all delivery nodes. The larger the mean, the more common the loading and unloading time exceeds the limit. The variance of the docking time deviation is the degree of dispersion of the docking deviation of each node. A large variance indicates that the loading and unloading time fluctuates drastically and the planning and prediction are inaccurate.

[0127] In this embodiment, the mean of the arrival time deviation is the average value of the arrival time deviation of each node. A positive mean and a large value indicate that the overall delivery is generally late. The variance of the arrival time deviation represents the magnitude of the delay time fluctuation at each point.

[0128] In this embodiment, the total length of the planned route is the total mileage of the final delivery route plan. The formula for calculating the route deviation rate is: cumulative length of the deviation segment / total length of the planned route; the value ranges from 0 to 1. The larger the value, the higher the proportion of vehicles deviating significantly from the planned route, and the worse the model route planning fit.

[0129] In this embodiment, the docking reliability score is calculated as a 0-1 range score by jointly calculating the mean and variance of the docking deviation; the lower the score, the greater the model's error in predicting loading and unloading docking time, and the worse the docking planning reliability. The timeliness reliability score is calculated as a 0-1 range score by jointly calculating the mean and variance of the arrival time deviation; the lower the score, the greater the model's deviation in predicting transit time, and the weaker its on-time delivery planning capability. The model confidence adjustment factor is a comprehensive correction coefficient obtained by weighting and summing the route execution deviation rate, docking reliability score, and timeliness reliability score with fixed weights, quantifying the overall planning deviation of the model.

[0130] In this embodiment, the moving average filter smooths the mean of the confidence adjustment factors from multiple consecutive delivery trips, eliminating random interference caused by sudden traffic jams or temporary loading / unloading delays, and avoiding frequent changes to model parameters due to a single anomaly. The smoothing adjustment is a stable correction value after the moving average filter process eliminates random fluctuations, serving as the basis for parameter adjustment decisions.

[0131] In this embodiment, the preset adjustment threshold is a tiered judgment threshold, which is divided into a first adjustment threshold and a second adjustment threshold. The first adjustment threshold is greater than the second adjustment threshold, which is used to classify three types of parameter adjustment conditions.

[0132] The first adjustment threshold is the critical value for determining severe deviations in the model's comprehensive planning. It is jointly calibrated by dividing the intervals using batch historical delivery data statistics and superimposing on-board real-time computing power constraints: the smoothing adjustment amount of all delivery tasks in the collection area for two consecutive months is collected, and the frequency distribution of all values ​​is statistically analyzed to distinguish three types of working conditions: low deviation, medium deviation, and severe deviation. All delivery batches with severe deviations, such as large-scale trajectory deviations, multi-node loading and unloading timeouts, and large-scale overall delivery delays, are screened out, and the minimum value of the smoothing adjustment amount of this type of batch is extracted as the initial value of the first adjustment threshold. At the same time, the single-time optimization computing power delay constraint of the on-board dispatch terminal is introduced. If the threshold value is too small, it will frequently trigger the expansion of the search width and the increase of candidate solutions, causing calculation timeouts. The initial value is adjusted upwards to constrain this. Then, extreme high-deviation delivery samples such as morning peak full-area congestion, summer high temperature multi-node delayed loading and unloading, and long-term road construction detours are used for retesting and verification. After fine-tuning the values, the first adjustment threshold is obtained, and it is ensured that this value is greater than the second adjustment threshold.

[0133] The second adjustment threshold is jointly calibrated using a large-scale statistical analysis of high-quality, low-disturbance delivery samples combined with constraints on onboard computing power: High-quality delivery records from two consecutive months that perfectly adhere to the planned route and show no significant deviations in stop and arrival times at each node are retrieved, and the smoothing adjustment amount corresponding to each delivery is extracted. The maximum smoothing adjustment amount of all high-quality deliveries is then used as the initial value of the second adjustment threshold. If the threshold is set too high, normal deliveries with minor deviations may be mistakenly classified as low-deviation deliveries, prematurely reducing the search width and leading to a decrease in subsequent path planning accuracy. If the threshold is set too low, the effect of shrinking parameters and saving onboard computing power cannot be fully utilized. Based on the balance constraints between accuracy and computing power, the initial value is slightly corrected. Multiple sets of retest samples are then selected for verification, including those from off-peak seasons in spring and autumn, low-congestion conditions at normal temperatures, and stable loading and unloading operations, ensuring that the smoothing adjustment amount of all high-quality deliveries is less than the threshold, while the smoothing adjustment amount of moderate-deviation delivery samples is greater than the threshold. Finally, the second adjustment threshold is obtained, ensuring that it is less than the first adjustment threshold.

[0134] In this embodiment, the first adjustment threshold is a high deviation critical value. A smoothing adjustment amount greater than or equal to this value indicates that the model planning error is serious and the search range needs to be widened. The second adjustment threshold is a low deviation critical value. A smoothing adjustment amount lower than this value indicates that the model planning accuracy is high and the search range can be narrowed to reduce computing power.

[0135] In this embodiment, the path search width is the range of alternative detour routes traversed by the adaptive genetic scheduling model during optimization. A larger width allows for the traversal of more alternative routes, but also increases computational overhead. The number of candidate solutions represents the number of alternative delivery routes computed synchronously in parallel.

[0136] In this embodiment, the fifth preset step size is a fixed increment for increasing the path search width and the number of candidate solutions in a single operation when the smoothing adjustment amount exceeds the first adjustment threshold. The sixth preset step size is a fixed decrease for decreasing the path search width and the number of candidate solutions in a single operation when the smoothing adjustment amount is less than the second adjustment threshold.

[0137] The fifth preset step size is determined by combining multiple sets of severe disturbance simulation conditions with the upper limit constraint of vehicle computing latency. This involves calibrating a severely skewed road network with large-scale congestion, high temperatures on multiple road sections, and loading / unloading delays at multiple nodes. With other model parameters fixed, multiple sets of step size increments are used to conduct batch path optimization tests. The completeness of detour branch traversal, the detection rate of the optimal loss route for fresh produce, and the calculation time for a single optimization are recorded for each set of step sizes. The core constraint is to fully traverse low-loss alternative routes and not overlook high-quality detour schemes under severe deviation scenarios, while ensuring that the time for a single planning operation does not exceed the upper limit of vehicle terminal latency. The initial increment is then obtained by retesting and verifying real high-deviation delivery samples from morning rush hour congestion and summer high temperatures combined with road construction. The corrected values ​​are then used to obtain the fifth preset step size.

[0138] The sixth preset step size is jointly calibrated using a low-disturbance, stable delivery simulation control test combined with a baseline constraint on path planning accuracy. A low-deviation simulated delivery network with smooth traffic flow, suitable temperature, and stable loading / unloading times is used, while keeping other basic parameters of the model unchanged. Multiple reduction step sizes are set at the gradient, and path optimization calculations are performed in batches. The time consumed by each optimization calculation, the accuracy of the optimal delivery route matching, and the node timeliness prediction error are collected for each step size. Initial reductions are obtained by using the dual constraints of minimizing onboard computing power consumption in low-deviation scenarios while not losing the optimal short-distance, smooth delivery routes. Then, real, high-quality delivery samples from spring and autumn off-peak seasons, suburban areas with low traffic flow, and standardized loading / unloading operations are retrieved for retesting and verification. After correcting the values, the sixth preset step size is obtained.

[0139] As can be seen from the above, this embodiment constructs a model confidence adjustment factor by integrating path deviation rate, docking and timeliness reliability scores. After smoothing with moving average filtering, it adaptively adjusts the path search width and the number of candidate solutions using thresholds. This solves the shortcomings of the original model parameters, which cannot be quantified and iterated, and the adjustment logic is singular. It comprehensively evaluates the model planning confidence level through multi-dimensional deviation indicators, and filters to eliminate parameter misadjustments caused by single abnormal delivery data. The optimization range is adjusted in a two-level threshold stepwise manner. When the deviation is large, the search range is widened to improve optimization fault tolerance; when the deviation is small, the search range is narrowed to improve computational efficiency. The parameters are kept stable in the stable interval, realizing adaptive and refined iteration of the scheduling model and improving the overall performance of dynamic cold chain path optimization.

[0140] Corresponding to the dynamic cold chain distribution strategy optimization method in the above embodiment, Figure 2This is a structural block diagram of a dynamic cold chain distribution strategy optimization system provided in one embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The dynamic cold chain distribution strategy optimization system 20 includes: a region determination module 21, a feature extraction module 22, a disturbance vector module 23, a state evaluation module 24, a target optimization module 25, a path reconstruction module 26, a multi-dimensional vector module 27, and an optimization scheme module 28.

[0141] Among them, the region determination module 21 is used to determine the service coverage of each cold chain delivery vehicle based on historical order data and real-time order data of the target delivery area by analyzing the spatiotemporal distribution of cold chain delivery vehicles and the quality decay characteristics of fresh goods, and generate an initial delivery route as a benchmark delivery route; the service coverage and the benchmark delivery route together constitute the first spatiotemporal constraint region. Feature extraction module 22 is used to extract features from the baseline delivery route to obtain the feature vector of the baseline delivery route; The disturbance vector module 23 is used to acquire road condition data, ambient temperature data and vehicle operation data within the first spatiotemporal constraint region, and to preprocess and extract features from the road condition data, ambient temperature data and vehicle operation data, and then fuse them to obtain the first disturbance feature vector. The state assessment module 24 is used to input the first disturbance feature vector into the preset delivery state assessment model to obtain the stability assessment result and disturbance level classification of the current delivery network. The target optimization module 25 is used to obtain the current delivery route of the cold chain delivery vehicle, and based on the stability assessment results and disturbance level classification, determine the target optimization route segment that needs to be replanned or partially adjusted from the current delivery route. The path reconstruction module 26 is used to partially reconstruct and adjust the target optimized path segment, generate the reconstructed path segment, splice the reconstructed path segment with the unadjusted part of the current delivery path and extract its features to generate the second delivery path feature vector, and input the second delivery path feature vector and the benchmark delivery path feature vector into the preset path deviation identification model to obtain the path deviation and delay risk identification results. The multidimensional vector module 27 is used to perform fusion analysis on the path offset and delay risk identification results and the first disturbance feature vector to extract multidimensional disturbance feature variables; The optimization module 28 is used to input multi-dimensional disturbance feature variables and baseline delivery path feature vectors into a preset adaptive path planning and delivery scheduling model to obtain the final delivery path scheme after optimization.

[0142] See Figure 3 , Figure 3This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the shown region determination module 21, feature extraction module 22, perturbation vector module 23, state evaluation module 24, target optimization module 25, path reconstruction module 26, multidimensional vector module 27, and optimization scheme module 28 are as follows.

[0143] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0144] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0145] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0146] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the dynamic cold chain distribution strategy optimization method provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0147] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0148] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0149] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0150] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0151] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces or units, or they may be electrical, mechanical, or other forms of connection.

[0152] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0153] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0154] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for optimizing dynamic cold chain distribution strategies, characterized in that, include: Based on historical and real-time order data of the target delivery area, the service coverage of each cold chain delivery vehicle is determined by analyzing the spatiotemporal distribution of cold chain delivery vehicles and the quality degradation characteristics of fresh goods, and an initial delivery route is generated as the benchmark delivery route; the service coverage and the benchmark delivery route together constitute the first spatiotemporal constraint area. Feature extraction is performed on the baseline delivery route to obtain the baseline delivery route feature vector; Within the first spatiotemporal constraint region, road condition data, ambient temperature data, and vehicle operation data are acquired, and the road condition data, ambient temperature data, and vehicle operation data are preprocessed and feature extracted, and then fused to obtain the first disturbance feature vector. The first disturbance feature vector is input into a preset delivery status assessment model to obtain the stability assessment result and disturbance level classification of the current delivery network. Obtain the current delivery route of the cold chain delivery vehicle, and based on the stability assessment results and disturbance level classification, determine the target optimized route segment that needs to be replanned or partially adjusted from the current delivery route; The target optimized path segment is partially reconstructed and adjusted to generate a reconstructed path segment. The reconstructed path segment is then spliced ​​with the unadjusted part of the current delivery path and its features are extracted to generate a second delivery path feature vector. The second delivery path feature vector and the baseline delivery path feature vector are then input into a preset path deviation identification model to obtain the path deviation and delay risk identification results. The path offset and delay risk identification results and the first disturbance feature vector are fused and analyzed to extract multi-dimensional disturbance feature variables; The multi-dimensional perturbation feature variables and the baseline delivery path feature vector are input into a preset adaptive path planning and delivery scheduling model to obtain the final delivery path scheme after optimization.

2. The dynamic cold chain distribution strategy optimization method according to claim 1, characterized in that, The process of obtaining the current delivery route of the cold chain delivery vehicle, and based on the stability assessment results and disturbance level classification, determining the target optimized route segment from the current delivery route that requires route replanning or local adjustment, includes: Based on the preset perturbation type-weight mapping table, corresponding fuzzy evaluation weights are assigned to different identified perturbation types. The fuzzy comprehensive evaluation method is used to evaluate each segment along the current delivery route based on multiple factors, and a disturbance risk score curve along the route is generated. The multiple factors include road congestion, temperature fluctuation, time window urgency, and vehicle remaining range. Based on the disturbance risk scoring curve, continuous road segments with scores greater than or equal to a preset risk scoring threshold are extracted to obtain high-risk path segments. The real-time traffic flow data and fresh produce quality decay rate corresponding to the high-risk path segment are obtained. The start and end points of the high-risk path segment are corrected based on the flow-decay coupling model to obtain the target optimized path segment.

3. The dynamic cold chain distribution strategy optimization method according to claim 2, characterized in that, The method employs fuzzy comprehensive evaluation to assess multiple factors along each segment of the current delivery route, generating a disturbance risk score curve along the route, including: The current delivery route is divided into several continuous segments according to the interval between adjacent stops, and the measured factor values ​​of each segment are obtained. Each evaluation factor for each road segment is divided into several fuzzy levels, including a first risk level, a second risk level, and a third risk level, wherein the first risk level is lower than the second risk level, and the second risk level is lower than the third risk level. Construct membership functions for each factor to each fuzzy level, wherein the membership functions adopt a trapezoidal distribution or a triangular distribution; The membership matrix corresponding to each road segment is calculated based on the measured factor values ​​of each road segment. The membership matrix is ​​then combined with the fuzzy evaluation weights to obtain the membership vector of each road segment to each risk level. Extract the membership value of the third risk level from the membership vector as the disturbance risk score of the corresponding road segment, and connect the disturbance risk scores of all road segments in the order of the paths to generate the disturbance risk score curve.

4. The dynamic cold chain distribution strategy optimization method according to claim 1, characterized in that, The process involves locally reconstructing and adjusting the target optimized path segment to generate a reconstructed path segment. This reconstructed path segment is then concatenated with the unadjusted portion of the current delivery path, and features are extracted to generate a second delivery path feature vector, including: Obtain the start and end node information of the target optimized path segment and the corresponding constraints. The constraints include the stop time window, vehicle remaining capacity, remaining freshness duration of fresh goods, and collected road condition data and ambient temperature data. With the optimization objectives of minimizing delivery time, minimizing cargo quality degradation, and maximizing path reliability, a local optimization algorithm is used to replan the feasible path between the start and end nodes, generating the reconstructed path segment. The reconstructed path segment is then joined with the unadjusted portion of the current delivery path to form an updated complete delivery path. Feature extraction is performed on the updated complete delivery route to obtain the second delivery route feature vector.

5. The dynamic cold chain distribution strategy optimization method according to claim 4, characterized in that, The local optimization algorithm employs particle swarm optimization; the process of replanning feasible paths between the start and end nodes using the local optimization algorithm to generate reconstructed path segments includes: Based on the road network topology and the number of stops between the start and end nodes of the target optimized path segment, the solution space of the particle swarm optimization algorithm is determined, where the position vector of each particle in the solution space represents a visiting order of intermediate stops between the start and end nodes. Based on the disturbance level classification and the road condition data in the first disturbance feature vector, the hyperparameter combination of the particle swarm optimization algorithm is determined.

6. The dynamic cold chain distribution strategy optimization method according to claim 5, characterized in that, The hyperparameters include inertia weights and acceleration factors; The step of determining the hyperparameter combination of the particle swarm optimization algorithm based on the disturbance level classification and the road condition data in the first disturbance feature vector includes: Obtain the disturbance level classification, which includes severe disturbances and minor disturbances; If the disturbance level is classified as a severe disturbance, the reference value of the inertial weight is increased based on a first preset step size, and the reference value of the acceleration factor is decreased based on a second preset step size. If the disturbance level is classified as a minor disturbance, the reference value of the inertial weight is reduced based on a third preset step size, and the reference value of the acceleration factor is increased based on a fourth preset step size.

7. The dynamic cold chain distribution strategy optimization method according to claim 1, characterized in that, After inputting the multi-dimensional perturbation feature variables and the baseline delivery path feature vector into a preset adaptive path planning and delivery scheduling model to obtain the optimized final delivery path scheme, the method further includes: The final delivery route plan is sent to the on-board terminal of the corresponding cold chain delivery vehicle and pushed to the monitoring dashboard of the delivery management platform. Collect actual operating data of the cold chain delivery vehicle during the delivery process according to the final delivery route plan. The actual operating data includes the actual driving trajectory, actual stopping time and actual arrival time of each node. The actual operating data is compared with the expected data in the final delivery route plan to generate a plan execution deviation report; Based on the deviation report of the proposed scheme, adjust the algorithm running parameters in the adaptive route planning and delivery scheduling model.

8. The dynamic cold chain distribution strategy optimization method according to claim 7, characterized in that, The step of comparing the actual operational data with the expected data in the final delivery route plan to generate a plan execution deviation report includes: Obtain the actual driving trajectory, actual stopping time, and actual arrival time of each node from the actual operation data; The actual driving trajectory is spatially aligned with the planned route in the final delivery route scheme. The deviation distance between the actual driving route and the planned route is calculated for each road segment. Road segments with a deviation distance greater than or equal to a preset deviation threshold are marked as trajectory deviation road segments. The actual stopping time is compared with the planned stopping time in the final delivery route plan, and the stopping time deviation value of each stopping point is calculated. The stopping time deviation value is the actual stopping time minus the planned stopping time. The actual arrival time of each node is compared with the planned arrival time in the final delivery route plan, and the arrival time deviation value of each node is calculated. The arrival time deviation value is the actual arrival time minus the planned arrival time. Based on the trajectory deviation segment, the stop time deviation value, and the arrival time deviation value, a plan execution deviation report is generated, including the deviation type, deviation location, deviation amount, and deviation duration.

9. The dynamic cold chain distribution strategy optimization method according to claim 8, characterized in that, The step of adjusting the algorithm operating parameters in the adaptive route planning and delivery scheduling model according to the deviation report based on the scheme includes: Extract the cumulative length of the trajectory deviation segment, the mean and variance of the stop time deviation value, and the mean and variance of the arrival time deviation value from the deviation report of the scheme execution; The path execution deviation rate is calculated based on the ratio of the cumulative length of the deviation segments to the total length of the planned path in the final delivery route scheme. Calculate the docking reliability score based on the mean and variance of the docking time deviation values; Calculate the timeliness reliability score based on the mean and variance of the arrival time deviation values; The path execution deviation rate, docking reliability score, and timeliness reliability score are weighted and fused to obtain the model confidence adjustment factor. The confidence adjustment factor of the model is smoothed using a moving average filter to obtain the smoothing adjustment amount. The smoothing adjustment amount is compared with a preset adjustment threshold, which includes a first adjustment threshold and a second adjustment threshold, wherein the first adjustment threshold is greater than the second adjustment threshold. If the smoothing adjustment amount is greater than or equal to the first adjustment threshold, then the path search width and the number of candidate solutions in the adaptive path planning and delivery scheduling model are increased based on the fifth preset step size. If the smoothing adjustment amount is less than the second adjustment threshold, then the path search width and the number of candidate solutions in the adaptive path planning and delivery scheduling model are reduced based on the sixth preset step size. If the smoothing adjustment amount is greater than or equal to the second adjustment threshold and less than the first adjustment threshold, then the current algorithm running parameters remain unchanged.

10. A dynamic cold chain distribution strategy optimization system, characterized in that, include: The region determination module is used to determine the service coverage of each cold chain delivery vehicle based on historical and real-time order data of the target delivery area by analyzing the spatiotemporal distribution of cold chain delivery vehicles and the quality degradation characteristics of fresh goods, and to generate an initial delivery route as a benchmark delivery route; the service coverage and the benchmark delivery route together constitute the first spatiotemporal constraint region. The feature extraction module is used to extract features from the baseline delivery route to obtain a feature vector of the baseline delivery route; The disturbance vector module is used to acquire road condition data, ambient temperature data and vehicle operation data within the first spatiotemporal constraint region, and to preprocess and extract features from the road condition data, ambient temperature data and vehicle operation data, and then fuse them to obtain the first disturbance feature vector. The status assessment module is used to input the first disturbance feature vector into a preset delivery status assessment model to obtain the stability assessment result and disturbance level classification of the current delivery network. The target optimization module is used to obtain the current delivery route of the cold chain delivery vehicle, and based on the stability assessment results and disturbance level classification, determine the target optimization route segment that needs to be replanned or partially adjusted from the current delivery route. The path reconstruction module is used to partially reconstruct and adjust the target optimized path segment to generate a reconstructed path segment. The reconstructed path segment is then spliced ​​with the unadjusted part of the current delivery path and its features are extracted to generate a second delivery path feature vector. The second delivery path feature vector and the baseline delivery path feature vector are then input into a preset path deviation identification model to obtain the path deviation and delay risk identification results. The multidimensional vector module is used to fuse and analyze the path offset and delay risk identification results and the first disturbance feature vector to extract multidimensional disturbance feature variables. The optimization module is used to input the multi-dimensional perturbation feature variables and the baseline delivery path feature vector into a preset adaptive path planning and delivery scheduling model to obtain the final delivery path scheme after optimization.

Images