An intelligent scheduling method for fresh food distribution based on performance constraints

Abstract

The application belongs to the technical field of logistics scheduling, and discloses a fresh food distribution intelligent scheduling method based on performance constraint, which combs the whole-link core performance nodes from order generation to user signing, builds four-dimensional index library of product characteristics, user demand and the like, realizes dynamic updating of indexes by associating fresh food life cycle with real-time environment data; meanwhile, a three-level management and control system is divided and the default cost is quantified, a digital twin model and a constraint conflict prediction mechanism are combined to identify node constraint conflicts in advance, hard constraint is one-vote veto, and risk constraint is pre-control; a weighted K-means clustering algorithm model is constructed based on constraint priority, order dynamic aggregation and grouping are realized by combining taboo rules, and default risk after merging is avoided; meanwhile, a digital image including three dimensions and a four-level rating system are built for transport capacity, transport capacity matching is realized through two-level verification of hard constraint one-vote veto and comprehensive matching degree weighted calculation, and a transport capacity dynamic adaptation mechanism is matched to replace abnormal transport capacity in time.

Description

Technical Field This invention belongs to the field of logistics scheduling technology, specifically a smart scheduling method for fresh food delivery based on fulfillment constraints. Background Technology Currently, the mainstream technical solutions for fresh food delivery scheduling in the industry all revolve around the vehicle routing problem (VRP), and generally adopt heuristic algorithms such as genetic algorithms and ant colony algorithms. The core optimization objectives are the shortest delivery route and the lowest operating cost. They only use single-dimensional indicators such as last-mile delivery time and vehicle temperature zone as hard constraints to meet the basic needs of fresh food delivery. However, the existing technologies still have the following technical problems: The existing solution only uses last-mile delivery time as a single constraint, which does not cover the entire fulfillment lifecycle of order sorting, pre-cooling, warehousing, transportation and last-mile signing. It also does not establish a differentiated and graded constraint system for fresh food categories, user needs and compliance requirements, which makes it easy to have partial compliance but full-chain fulfillment breaches. The order aggregation and capacity matching did not take the priority, compatibility and default risk of performance constraints as the core matching criteria, resulting in constraint conflicts and mismatch between capacity and order performance requirements after order merging, which significantly increased the probability of performance default and the loss rate of fresh produce. The optimization objective is too singular. It fails to quantify and incorporate the costs of breach of contract, fresh produce spoilage, and user experience loss into the optimization function, resulting in optimization results that can only achieve local cost optimization and cannot guarantee the optimal overall performance. Most existing solutions are static, one-time scheduling solutions, lacking real-time perception and graded response to disturbances at all fulfillment nodes across the entire chain. They also lack a pre-emptive warning mechanism for default risks, making it impossible to avoid fulfillment risks in advance, resulting in a persistently high default rate. Furthermore, they fail to consider the life cycle characteristics of fresh produce and user fulfillment preferences, lacking adaptive adjustment capabilities and making it difficult to adapt to the dynamic changes in fresh produce delivery scenarios. Summary of the Invention The purpose of this invention is to provide an intelligent scheduling method for fresh food delivery based on fulfillment constraints, so as to solve the problems mentioned in the background art.

[0001] To achieve the above objectives, the present invention provides the following technical solution: an intelligent scheduling method for fresh food delivery based on fulfillment constraints, comprising the following specific steps: Preferably, the constraint modeling stage sorts out the core fulfillment nodes of the entire chain of fresh food delivery from order generation, sorting and packaging, pre-cooling treatment, outbound verification, urban distribution transportation, last-mile delivery to user signature, and builds a fulfillment constraint indicator library for each node. The indicators include four dimensions: category characteristics, user needs, compliance requirements, and transportation capacity limits. Each indicator is associated with fresh food category life cycle parameters and real-time environmental data to realize dynamic updates of constraint indicators. The lifecycle parameters for fresh produce are mainly related to core data such as shelf life, respiration intensity, and temperature and humidity tolerance. Real-time environmental data includes data on environmental temperature and humidity, atmospheric pressure, and road congestion in the delivery area. The two are linked through data association mapping rules. When the environmental temperature and humidity exceed the tolerance of fresh produce, the delivery lifecycle threshold for that category is automatically shortened. When the road congestion level increases, the performance constraint indicators of urban distribution transportation nodes are adjusted simultaneously. All association rules are embedded in the digital twin model of performance constraints to achieve automatic and dynamic updates of indicators.

[0002] Based on the inviolability of constraints and the degree of impact of breach of contract, all constraints are divided into a three-level control system and the cost of breach of contract and trigger threshold of each level of constraints are quantified. Hard constraints include mandatory temperature zone requirements for fresh products, food safety compliance requirements, and requirements for complete traceability information. Soft constraints include user-adjustable signing time and delivery order preferences. Risk constraints are types of constraints based on pre-control measures such as fresh product spoilage rate thresholds and transportation capacity fatigue thresholds. Construct a digital twin model of performance constraints that integrates real-time data perception, integrates IoT data interfaces, and realizes digital presentation, unified management and control, full-process calling and dynamic updating of constraints across the entire chain. At the same time, establish a constraint conflict prediction mechanism to identify potential conflicts between constraints at different nodes in advance.

[0003] The IoT data interface supports three connection methods: serial port, Ethernet, and 4G / 5G wireless communication, respectively adapting to data transmission of cold chain vehicle-mounted equipment, warehouse fixed equipment, and mobile delivery equipment. The interface adopts a standardized data protocol to achieve seamless connection with temperature control sensors, GPS positioning terminals, and electronic tag traceability equipment. Data transmission adopts an encrypted format to ensure the security and accuracy of real-time data. The interface supports multiple devices to connect simultaneously, and the maximum number of devices that can be connected to a single interface meets the daily needs of full-chain monitoring of delivery.

[0004] Preferably, the aggregation and matching stage is based on the full-link performance constraint indicator library, constraint hierarchical control system and constraint conflict prediction results built in the constraint modeling stage. It obtains the fresh food order pool to be delivered in real time, extracts the full-link performance constraint parameters, constraint level and priority weight corresponding to each order, integrates user performance preference profile and compliance priority, and constructs a weighted K-means clustering algorithm model that integrates constraint priority, user preference and compliance weight. The weighted K-means clustering algorithm model includes constraint compatibility verification factor and default risk prediction function. The core order of compliance priority set by this invention is food safety compliance > cold chain temperature control compliance > delivery timeliness compliance > user experience compliance. The priority can be fine-tuned for different fresh food categories. For cold chain sensitive categories such as seafood and meat, the priority of cold chain temperature control compliance is increased, while for fruits and vegetables and prepared meals, food safety compliance is kept first. The ranking result of compliance priority is directly converted into the weight coefficient of the corresponding algorithm and incorporated into the calculation logic of order aggregation.

[0005] The order pool is dynamically aggregated and grouped to generate several order delivery groups with optimal internal constraint compatibility, lowest risk of default, and highest user preference matching. At the same time, a unified performance constraint control standard and priority are set for each order delivery group. Order merging taboo rules are introduced. If the overall default probability of multiple orders exceeds a preset threshold after merging, or violates the compliance priority ranking, or conflicts with user preferences, the aggregation operation is prohibited. When executing the rules for order aggregation, compliance priority verification is performed first, followed by default probability calculation, and finally user preference matching verification. The three verifications are performed in this order. If any verification fails, the aggregation operation is terminated immediately. The verification results generate a dedicated verification report, recording the reasons for failure and relevant indicator data. The report is also stored in the order aggregation archive to provide data reference for subsequent order segmentation optimization.

[0006] Establish a dynamic adjustment mechanism for order aggregation, receive information such as new orders, order cancellations, and changes in user needs in real time, and dynamically split and reorganize the aggregated order groups.

[0007] Preferably, the capacity matching stage is based on the order delivery group generated in the aggregation matching stage and the set performance constraint control standards. For the cold chain delivery capacity resource pool, a capacity digital profile system with performance capability as the core is constructed. The capacity digital profile system includes three dimensions: physical attributes, historical performance attributes and real-time status. At the same time, a capacity performance capability rating system is introduced, and the rating results are dynamically correlated with historical performance data and real-time status. For each order delivery group generated during the aggregation and matching phase, a two-level matching verification is performed. The first level is a hard constraint veto verification, which directly filters out the capacity that does not meet the core hard constraints of the order group. The second level calculates the comprehensive matching degree between the remaining capacity and the order delivery group. Establish a dynamic capacity adaptation mechanism to monitor the capacity operation status in real time. If the capacity is abnormal, the matching process of backup capacity will be initiated.

[0008] Preferably, the path planning stage constructs a multi-objective performance constraint game optimization function based on the full-link node constraints of the constraint modeling stage, the order group constraints of the aggregation matching stage, and the capacity constraints and matching results of the capacity matching stage. The core optimization objective is to minimize the default risk of the performance constraints of the order delivery group, while taking into account four auxiliary optimization objectives: the lowest total delivery cost, the lowest fresh produce loss, the best delivery time, and the best user experience. The weight allocation of each objective is clearly defined. All constraints from the constraint modeling stage (full-link node constraints), the aggregation and matching stage (order group constraints), and the capacity matching stage (capacity constraints) are incorporated into the constraints of the optimization function. Hard constraints are used as boundary conditions, while soft constraints and risk constraints are used as dynamic adjustment factors. The NSGA-III multi-objective optimization algorithm with constraint priority is adopted. The NSGA-III multi-objective optimization algorithm integrates the Nash equilibrium concept from game theory to construct a three-party game model involving the platform, capacity, and users, and solves it to obtain the Pareto optimal solution set. Based on the order group's fulfillment constraint priority, capacity rating, and user preferences, the globally optimal delivery route plan is selected. At the same time, the time control threshold, operation standards, and abnormal triggering conditions for each fulfillment node are clearly defined. Real-time traffic prediction models and fresh produce spoilage prediction models are added to the route planning, and the route and node time thresholds are dynamically adjusted based on the prediction models.

[0009] Preferably, the disturbance response phase is based on the time control threshold, operation standards and abnormal triggering conditions of the fulfillment nodes specified in the path planning phase, and integrates real-time data from multiple systems such as IoT temperature control equipment, GPS positioning system, WMS warehousing system, TMS transportation management system and user feedback system. It collects the status data of the fulfillment nodes in the entire chain in real time, including sorting progress, pre-cooling temperature, real-time vehicle location, cold chain equipment operating parameters, user acceptance intention, road traffic conditions and other related data. The collected real-time data is compared with the performance node control thresholds set in the path planning stage to identify performance disturbance events. Based on the degree, scope and duration of the impact of the disturbance on the performance constraints, the performance disturbance events are divided into three levels: Level 1 disturbance is a major disturbance that is about to break through the hard constraint, Level 2 disturbance is a moderate disturbance that breaks through the threshold of the soft constraint, and Level 3 disturbance is a minor disturbance that does not affect the core performance. The identification of Level 3 disturbances is achieved by comparing real-time data from all nodes in the supply chain with preset thresholds. The system automatically triggers the identification results without manual intervention. After identification, a disturbance details sheet is generated immediately, which includes the node where the disturbance occurred, the type of disturbance, the scope of impact, and the estimated duration. The details sheet is simultaneously pushed to the corresponding node's operations and scheduling personnel, providing accurate information support for the execution of the tiered response strategy.

[0010] For different levels of disturbances, corresponding graded response strategies are triggered. Level 1 disturbances initiate emergency handling procedures, Level 2 disturbances initiate dynamic adjustment procedures, and Level 3 disturbances initiate early warning and alert procedures. Under the premise of not violating core performance constraints and balancing the interests of all parties, the additional costs incurred by scheduling and adjustment are controlled to the greatest extent possible. At the same time, a disturbance response effect evaluation mechanism is established to track the adjustment effect in real time.

[0011] Preferably, the risk warning stage is based on the real-time performance data, disturbance identification results and response effect evaluation data collected in the disturbance response stage, combined with historical performance big data, fresh food category characteristic data and environmental prediction data, to construct a performance default risk prediction model based on LSTM long short-term memory network and reinforcement learning. The performance default risk prediction model predicts the full-chain performance default probability of each order and each order delivery group in real time. The model input dimensions include various relevant indicators such as performance progress deviation, road traffic prediction data, cold chain equipment operation status, fresh food loss trend, and user feedback information. A multi-level risk threshold is set. When the predicted probability of default exceeds the preset risk threshold, a tiered early warning mechanism is triggered, and multiple risk handling plans are generated simultaneously. These plans include various methods such as capacity adjustment, route optimization, user communication, and compliance remediation. Based on the principles of minimizing handling costs, maximizing performance guarantee, and minimizing the impact on user experience, a multi-objective decision-making algorithm is used to select the optimal handling plan and execute it automatically. At the same time, the execution effect of the handling plan is tracked in real time. If the probability of default is still higher than the threshold after handling, an upgraded handling process is triggered.

[0012] Preferably, during the iterative optimization phase, after each delivery task is completed, the closed-loop management results and handling data from the risk warning phase are combined to collect full-chain actual performance data, including the completion status of performance nodes, the execution status of constraints, the actual loss rate of fresh produce, the performance of transportation capacity, user satisfaction, the cost of handling defaults, and other relevant data, to establish a performance data knowledge base. The actual performance data is compared and analyzed with the preset performance constraint standards and optimization objectives to identify optimization points in the constraint model, order aggregation algorithm, capacity matching rules, route planning algorithm, and risk warning model, and to identify performance differences for different fresh food categories and delivery scenarios. Through reinforcement learning algorithms, various relevant parameters such as the threshold and weight of the performance constraint model, the parameters of the order aggregation algorithm, the weight allocation of capacity matching, the target weight of route planning, and the threshold of risk warning are adaptively and iteratively optimized. The optimized model parameters are synchronously updated to the performance constraint digital twin model in the constraint modeling stage, the order aggregation algorithm in the aggregation matching stage, the path planning algorithm in the path planning stage, and the early warning model in the risk warning stage; a scenario adaptation module is established to automatically call the optimized parameters and models for different fresh food categories and different delivery scenarios.

[0013] The beneficial effects of this invention are as follows: 1. This invention sorts out the core fulfillment nodes of the entire chain from order generation to user signature, builds a four-dimensional indicator library including category characteristics, user needs, etc., and links fresh food life cycle with real-time environmental data to realize dynamic updates of indicators; at the same time, it divides a three-level control system and quantifies the cost of breach of contract, and combines a digital twin model and constraint conflict prediction mechanism to identify node constraint conflicts in advance, with hard constraints having a veto and risk constraints being managed in advance, realizing unified digital control of constraints throughout the entire chain, improving the fulfillment compliance of the entire fresh food delivery process, and adapting to the differentiated fulfillment needs of different fresh food categories.

[0014] 2. This invention constructs a weighted K-means clustering algorithm model based on constraint priority, and combines tabu rules to achieve dynamic aggregation and grouping of orders, avoiding the risk of default after merging. At the same time, it builds a digital profile of transportation capacity with three dimensions and a four-level rating system. Through a two-level verification of hard constraint veto and comprehensive matching degree weighted calculation, it achieves accurate matching of transportation capacity, and with a dynamic transportation capacity adaptation mechanism, it replaces abnormal transportation capacity in a timely manner. It ensures optimal constraint compatibility within the order group, realizes the rational allocation of transportation capacity resources, matches high-risk orders with high-rated transportation capacity, and adapts low-risk orders to low-rated transportation capacity, reducing the probability of performance default and fresh food loss rate, and improving the utilization rate of logistics resources.

[0015] 3. This invention constructs a multi-objective game optimization function with the core objective of minimizing performance default risk, and combines it with a prediction model to achieve dynamic path planning; it integrates data from multiple systems to identify performance disturbances and classifies them into three levels of response, triggering differentiated handling strategies for different disturbances; it constructs a risk prediction model based on LSTM and reinforcement learning to achieve graded early warning and optimal contingency plan selection; after delivery is completed, it collects full-link data to build a knowledge base, iteratively optimizes the parameters of each model through reinforcement learning, and combines it with a scenario adaptation module to achieve adaptive model invocation for different categories and scenarios; it realizes the upgrade from static scheduling to dynamic adaptive scheduling, continuously reducing the performance default rate, while balancing delivery costs, fresh produce spoilage, and user experience. Attached Figure Description Figure 1 This is a flowchart of the intelligent dispatching process for fresh food delivery in this invention; Figure 2 This is a flowchart of the performance constraint analysis and conflict prediction process of this invention. Detailed Implementation The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] like Figures 1 to 2 As shown, this embodiment of the invention provides an intelligent scheduling method for fresh food delivery based on fulfillment constraints, including the following specific steps: The constraint modeling phase involves identifying key fulfillment nodes across the entire fresh food delivery chain, from order generation, sorting and packaging, pre-cooling, outbound verification, urban distribution, last-mile delivery to customer signature. For each node, a fulfillment constraint indicator library is constructed, encompassing four dimensions: category characteristics, user needs, compliance requirements, and transport capacity limits. Each indicator is linked to fresh food category lifecycle parameters and real-time environmental data, such as ambient temperature and traffic congestion index, enabling dynamic updates of the constraint indicators. Category characteristic indicators include fresh food shelf life, respiration rate, and temperature and humidity sensitivity thresholds. User need indicators include flexible delivery times, door-to-door priority, and special delivery requirements. Compliance requirement indicators include food safety traceability, cold chain temperature control compliance, and delivery timeliness compliance. Transport capacity limit indicators include vehicle range, temperature zone adjustment range, and maximum load capacity. The entire supply chain is divided into seven core nodes: order generation, sorting and packaging, pre-cooling, outbound verification, urban distribution transportation, last-mile delivery, and customer signature. Each node is configured with a separate subset of performance constraint indicators. Category characteristic indicators are associated with the pre-cooling and urban distribution transportation nodes, user demand indicators are associated with the last-mile delivery and customer signature nodes, compliance requirement indicators cover all nodes, and capacity limit indicators are associated with the urban distribution transportation and last-mile delivery nodes. The indicator data of each node is collected and updated in real time at a frequency of once every minute.

[0017] Based on the inviolability of constraints and the degree of impact of breach of contract, all constraints are divided into a three-level control system, and the cost of breach and trigger threshold of each level of constraints are quantified. Hard constraints (veto type) include unbreakable content such as mandatory temperature zone requirements for fresh products, food safety compliance requirements, and complete traceability information requirements. Soft constraints (dynamically compatible type) include compatible content such as user-adjustable signing time and delivery order preferences. Risk constraints (early warning type) are constraint types based on the threshold of fresh product loss rate and the fatigue threshold of delivery capacity performance as the basis for early control. Hard constraint default cost accounting covers dimensions such as compensation for fresh produce spoilage, compliance penalties, and brand reputation loss. Risk constraint default cost accounting includes fresh produce spoilage costs and transportation capacity scheduling adjustment costs. Soft constraint default cost accounting covers user compensation costs and communication costs. Different fresh produce categories correspond to different default cost base values. Fruits and vegetables, meat, and seafood have base values ​​that are increased in a tiered manner. Trigger thresholds are dynamically set according to the life cycle characteristics of fresh produce categories.

[0018] The base value gradient for default costs is as follows: fruits and vegetables are based on a base value of 1, meat on a base value of 1.5, and seafood on a base value of 2. The default costs for risk constraints / soft constraints are converted based on the threshold breach coefficient on the base value.

[0019] Construct a digital twin model of performance constraints that integrates real-time data perception, integrates IoT data interfaces, and realizes digital presentation, unified management and control, full-process calling and dynamic updating of constraints across the entire chain. At the same time, establish a constraint conflict prediction mechanism to identify potential conflicts between constraints at different nodes in advance.

[0020] The digital twin model for performance constraints comprises four core layers: a data perception layer, a model mapping layer, a constraint control layer, and an interactive application layer. The data perception layer connects to IoT temperature control, GPS, WMS, TMS and other systems to achieve real-time data collection. The model mapping layer restores the performance scenarios and constraint relationships of the entire fresh food delivery chain in a 1:1 ratio. The constraint control layer realizes the dynamic updating of constraint indicators and conflict prediction. The interactive application layer provides a visual operation interface for dispatchers. The time synchronization error between model data updates and actual performance scenarios is controlled within 1 minute.

[0021] The constraint conflict prediction mechanism performs cross-validation on the constraint indicators of the entire link performance nodes every 5 minutes. It first verifies the compatibility of different constraints within a single node, and then verifies the connectivity of constraints across nodes. The core criteria for determining a conflict are that the constraint indicator value exceeds the preset threshold and that different constraint requirements are mutually exclusive. After a conflict is identified, it is classified into three categories according to the degree of impact of the breach: major, moderate, and minor. Major conflicts immediately trigger hard constraint control strategies, while moderate and minor conflicts are simultaneously pushed to the scheduling end and optimization suggestions are given.

[0022] The aggregation and matching phase, based on the full-link performance constraint indicator library, constraint hierarchical management system, and constraint conflict prediction results built in the constraint modeling phase, acquires the fresh food order pool to be delivered in real time, extracts the full-link performance constraint parameters, constraint level, and priority weights corresponding to each order, with the weight coefficient positively correlated with the constraint level, and hard constraints having the highest weight. It incorporates user performance preference profiles and compliance priorities, prioritizing food safety compliance > timeliness compliance > experience compliance, and constructs a weighted K-means clustering algorithm model that integrates constraint priority, user preferences, and compliance weights. This model includes a constraint compatibility verification factor and a default risk prediction function. The user performance preference profile is based on historical order data mining to extract user preference weights for delivery timeliness, pick-up time, and packaging requirements. The weighted K-means clustering algorithm model takes as input an order full-link performance constraint parameter vector, a constraint priority weight matrix, and a user performance preference feature vector. The output includes order aggregation and clustering results, the estimated default risk of each order group, and constraint compatibility. The constraint compatibility verification factor ranges from [0, 1], and a factor value ≥ 0.8 indicates constraint compatibility. The default risk prediction function is F(merging risk) = ∑(number of single constraint conflicts × corresponding default cost coefficient) / total number of constraints in the order group. The clustering iteration terminates when the number of iterations ≥ 50 or the sum of squared errors within the cluster converges to 10. -3the following.

[0023] The user fulfillment preference profile is built based on the user's historical delivery order data over the past 90 days. It focuses on mining four core features: delivery time preference, door-to-door delivery requirements, packaging requirements, and delivery order preference. Preference weights are assigned to each feature and visualized. The profile is dynamically updated on a monthly basis. If a user raises new personalized requirements during a delivery, the profile is entered in real time and the corresponding feature weights are adjusted. At the same time, high-frequency preferences and temporary preferences are marked, and high-frequency preferences are prioritized when aggregating and grouping users.

[0024] The order pool is dynamically aggregated and grouped to generate several order delivery groups with optimal internal constraint compatibility, lowest risk of default, and highest user preference matching. At the same time, a unified performance constraint control standard and priority are set for each order delivery group. Order merging taboo rules are introduced. If the overall default probability of multiple orders exceeds a preset threshold after merging, or violates the compliance priority ranking, or conflicts with user preferences, the aggregation operation of such orders is prohibited to avoid performance risks. Establish a dynamic adjustment mechanism for order aggregation, receive information such as new orders, order cancellations, and changes in user needs in real time, and dynamically split and reorganize the aggregated order groups.

[0025] The conditions for triggering adjustments for new orders are that the constraint compatibility between a single new order and the existing order group is ≥0.8; the conditions for triggering adjustments for order cancellations are that the cancelled orders account for ≥30% of the total number of orders in the existing order group; and the conditions for triggering adjustments for changes in user needs are that the changes involve hard constraints or risk constraints. Splitting and reorganizing prioritizes ensuring the constraint compatibility and low default risk of the remaining order group. The adjustment process is completed within 1 minute and synchronized to the capacity matching stage.

[0026] The unified performance constraint control standard for order delivery groups takes the highest level of hard constraints within the group as the core benchmark, and simultaneously incorporates the risk constraint thresholds of most orders in the group and the preferences of high-frequency users. The core includes four key elements: unified control value for cold chain temperature zone, maximum delivery time limit, maximum fresh produce loss rate threshold, and user signature time adaptation range. After the control standard is set, a unique control code is generated simultaneously, which runs through the entire process of subsequent capacity matching and route planning, and serves as the core basis for performance control at each stage.

[0027] The capacity matching phase, based on the order delivery groups generated in the aggregation matching phase and the established performance constraint control standards, constructs a digital capacity profile system centered on performance capability for the cold chain delivery capacity resource pool. This digital capacity profile system includes three dimensions: physical attributes, historical performance attributes, and real-time status. It also introduces a capacity performance capability rating system (S / A / B / C four levels), with the rating results dynamically linked to historical performance data and real-time status. Physical attributes include temperature zone configuration, rated load, cold chain equipment precision, and endurance. Historical performance attributes include historical on-time rate, temperature control compliance rate, default handling efficiency, and loss control capability. Real-time status includes current location, remaining performance capacity, cold chain equipment operating status, and driver fatigue level. The core monitoring indicators for cold chain equipment operating status include five categories: equipment startup status, real-time temperature and humidity inside the warehouse, temperature and humidity fluctuation range, refrigeration system operating power, and equipment fault alarm signals. Each indicator has a set normal operation threshold; exceeding the threshold indicates abnormal cold chain equipment operation. Abnormal indicator data is pushed to the capacity dispatching terminal in real time, triggering equipment abnormality warnings and reminding relevant personnel to promptly verify and handle the situation.

[0028] The capacity fulfillment capability is quantified and scored on a 100-point scale, divided into four levels: S, A, B, and C. Level S is 90-100 points, Level A is 80-89 points, Level B is 60-79 points, and Level C is 0-59 points. The scoring indicators include historical fulfillment attributes, real-time status, and physical attributes. The actual values ​​of each indicator are converted to a 100-point scale and then weighted to calculate the final rating score. Historical fulfillment attributes account for 60% of the weight, including on-time rate (30%), temperature control compliance rate (20%), and loss control capability (10%). Real-time status accounts for 25% of the weight, including cold chain equipment operating status (10%), driver fatigue level (10%), and remaining carrying capacity (5%). Physical attributes account for 15% of the weight, including temperature zone configuration (6%), rated load (5%), and range (4%).

[0029] The physical attributes of the transport capacity are static basic data, which are manually updated only when the transport capacity equipment is upgraded or modified after the initial entry; the historical performance attributes are updated cumulatively on a daily basis, and the performance data of the previous day is automatically summarized and added to the profile at midnight every day; the real-time status attributes are dynamic data, which are collected and updated once every minute through vehicle terminals and IoT devices to ensure that the profile data can accurately reflect the real-time performance capability of the transport capacity. All attribute data are stored in the dedicated data archive of the transport capacity and support full-process traceability.

[0030] For each order delivery group generated during the aggregation and matching phase, a two-level matching verification is performed. The first level is a hard constraint veto verification, which directly filters out the capacity that does not meet the core hard constraints of the order group (such as temperature zone requirements and compliance requirements). The second level calculates the comprehensive matching degree between the remaining capacity and the order delivery group. The comprehensive matching degree is calculated by weighting constraint compatibility × 60%, capacity rating × 25%, and user preference adaptability × 15%. Among them, order groups with high default risk are given priority to matching with capacity with strong historical anomaly handling capabilities and a rating of S / A, while low-risk order groups can be matched with B / C level capacity to optimize costs. The constraint compatibility is directly adopted using the actual value of the constraint compatibility verification factor. The capacity rating is converted into numerical values ​​according to S level 1.0, A level 0.8, B level 0.6 and C level 0.4. The user preference fit is calculated by the number of matching items for actual user needs / the total number of user needs. After the three indicators are converted into numerical values, the comprehensive fit is calculated by weighting them. The value range is [0, 1]. A comprehensive fit of ≥0.7 is judged as a fit between capacity and order group.

[0031] The hard constraint veto verification checks each of the core hard constraints of the order group. The verification content mainly includes three core items: cold chain temperature zone configuration, food safety compliance qualifications, and traceability information uploading capability. If any item fails to meet the requirements of the order group, it is judged as a failure and directly excluded from the capacity matching range. The verification results are recorded in the capacity performance file in real time. The capacity that fails the verification three times in a row will be temporarily removed from the capacity resource pool and will be reinstated after rectification. After rectification, it must pass the hard constraint veto verification three times in a row, and the core hard constraint indicators such as cold chain equipment and compliance qualifications must meet 100% of the requirements before it can be reinstated into the capacity resource pool.

[0032] Establish a dynamic capacity matching mechanism to monitor the operation status of capacity in real time. If there are any abnormalities in capacity, such as cold chain failures or traffic delays, the matching process for backup capacity will be initiated.

[0033] The criteria for determining abnormal transport capacity include four categories: cold chain equipment failure, vehicle breakdown, driver fatigue, and timeout warnings caused by traffic control. All types of abnormalities are automatically identified through real-time status data. When matching backup transport capacity, priority is given to using transport capacity with the same rating, region, and available status. If there is insufficient transport capacity with the same rating, higher-rated transport capacity will be used. The matching process must meet the core hard constraints of the order group, and the overall matching response time shall not exceed 3 minutes.

[0034] The path planning stage is based on the full-link node constraints of the constraint modeling stage, the order group constraints of the aggregation and matching stage, and the capacity constraints and matching results of the capacity matching stage. It constructs a multi-objective performance constraint game optimization function with the core optimization objective of minimizing the default risk of the performance constraint of the order delivery group as the core optimization objective, while taking into account four auxiliary optimization objectives: the lowest total delivery cost, the lowest fresh produce loss, the best delivery time, and the best user experience. The weight allocation of each objective is clearly defined, with the core objective having a weight of ≥50%, and the weight of the auxiliary objectives being dynamically adjusted according to the order type. The core objective of minimizing the risk of breach of contract is weighted at a fixed 50%. The weights of the four auxiliary objectives are dynamically allocated according to the type of fresh food order. For high-priced cold chain categories such as seafood and meat, the weight of minimizing fresh food spoilage is increased. For supermarket bulk delivery orders, the weight of minimizing total delivery cost is increased. For community instant delivery orders, the weight of optimal delivery time and optimal user experience is increased. The total weight of all auxiliary objectives after adjustment is 50%, and the weight of a single auxiliary objective shall not exceed 20%. The weight adjustment rules are automatically triggered by the system based on the order category tags.

[0035] The weights of auxiliary objectives are allocated according to the principle of prioritizing the core needs of the product category. For example, for seafood / meat orders, the weight of minimizing freshness loss is set at 20%, and the remaining 30% is evenly distributed among the other three auxiliary objectives at 10%. For community instant delivery orders, the weights of optimal delivery time and optimal user experience are each set at 20%, and the remaining 10% is allocated to minimizing total delivery cost.

[0036] All constraints from the constraint modeling stage (full-link node constraints), the aggregation and matching stage (order group constraints), and the capacity matching stage (capacity constraints) are incorporated into the constraints of the optimization function. Hard constraints are treated as insurmountable boundary conditions, while soft constraints and risk constraints are used as dynamic adjustment factors. The NSGA-III multi-objective optimization algorithm with constraint priority is adopted. This algorithm integrates the Nash equilibrium concept from game theory to construct a three-party game model involving the platform, capacity, and users. It balances the interests of the three parties, avoids global performance imbalance caused by the optimal interests of a single entity, and obtains the Pareto optimal solution set. The three-party game model constructs the revenue functions of the platform, transportation capacity, and users. The platform's revenue F(P) = total delivery revenue - fulfillment and breach of contract costs - fresh produce spoilage costs; the transportation capacity's revenue F(Transport) = delivery service fees - vehicle operating costs; and the user's revenue F(Use) = user experience quantitative score - delivery waiting time costs. The NSGA-Ⅲ algorithm is set with a population size of 100, a maximum number of iterations of 80, a crossover probability of 0.85, and a mutation probability of 0.05. The Pareto optimal solution set is sorted in ascending order by the default risk value. The top 20% of solutions are selected and then the global optimal delivery route is selected in descending order by the total delivery cost.

[0037] Based on the order group's fulfillment constraint priority, transportation capacity rating, and user preferences, the globally optimal delivery route plan is selected. At the same time, the time control threshold, operation standards, and abnormal triggering conditions for each fulfillment node are defined, realizing the collaborative planning of time and route for the entire chain from outbound to receipt. Real-time traffic prediction and fresh produce spoilage prediction models are added to the route planning, and the route and node time thresholds are dynamically adjusted based on the above two prediction models.

[0038] The fresh produce loss prediction model combines the characteristics of fresh produce categories, real-time cold chain temperature and humidity, delivery time prediction data, and ambient temperature data to predict loss trends. It increases the prediction frequency for perishable fruits, vegetables, and seafood, and outputs the estimated status of fresh produce loss for each delivery segment in real time. The real-time traffic prediction model connects to the city's traffic big data platform, covering traffic information for the entire delivery route, accurately predicting congested sections, congestion duration, and alternative routes. The prediction results of both models are synchronized to the path planning algorithm in real time, providing data support for adjusting path and node thresholds.

[0039] The threshold for fulfillment node control is set differently according to the life cycle characteristics of fresh products. The threshold for pre-cooling treatment and urban distribution transportation nodes of perishable products is set to a narrower range, while the threshold for nodes of storable products is set to a wider range. At the same time, it is dynamically adjusted in combination with delivery distance, real-time road conditions and transportation capacity status. If the delivery time is expected to increase, the time threshold for the last delivery node is appropriately relaxed without breaking the hard constraints. After the node threshold is adjusted, it is updated to the full-chain fulfillment control system.

[0040] The disturbance response phase is based on the time control thresholds, operating standards and abnormal triggering conditions of the fulfillment nodes defined in the path planning phase. It integrates real-time data from multiple systems such as IoT temperature control equipment, GPS positioning system, WMS warehousing system, TMS transportation management system and user feedback system. It collects status data of the fulfillment nodes in the entire chain in real time, including sorting progress, pre-cooling temperature, real-time vehicle location, cold chain equipment operating parameters, user acceptance intention, road traffic conditions and other related data, to achieve real-time data synchronization and abnormal perception. The core data collection dimensions of the user feedback system include five categories: real-time status of user acceptance intention, satisfaction with delivery time, feedback on the appearance of fresh produce, evaluation of packaging integrity, and evaluation of delivery personnel service. It also supports users to actively submit feedback on delivery anomalies. All collected data is synchronized in real time. Among them, user acceptance intention and feedback on delivery anomalies are the core disturbance identification data. Once relevant information is collected, the system immediately triggers node data verification to assist in the accurate identification of performance disturbances.

[0041] The collected real-time data is compared with the fulfillment node control thresholds set in the route planning stage to identify fulfillment disturbance events. Based on the degree, scope, and duration of the disturbance's impact on fulfillment constraints, fulfillment disturbance events are divided into three levels: Level 1 disturbances are major disturbances where hard constraints are about to be breached, such as cold chain failures, excessive fresh produce spoilage, or compliance indicators about to exceed thresholds; Level 2 disturbances are moderate disturbances where soft constraints exceed thresholds, such as delivery delays or changes in user demand; and Level 3 disturbances are minor disturbances that do not affect core fulfillment, such as temporary traffic congestion or slight delays in the sorting process. Different levels of disturbances trigger corresponding tiered response strategies. Level 1 disturbances initiate emergency handling procedures, such as changing transport capacity, emergency replenishment, and activating alternative routes. Level 2 disturbances initiate dynamic adjustment procedures, such as adjusting delivery order and negotiating user receipt times. Level 3 disturbances initiate early warning and reminder procedures, such as reminding drivers and synchronizing user progress. Under the premise of not violating core performance constraints and balancing the interests of all parties, the additional costs incurred by scheduling adjustments are controlled to the greatest extent possible. At the same time, a disturbance response effectiveness evaluation mechanism is established to track the adjustment effects in real time and ensure the effectiveness of the response strategy.

[0042] The core evaluation indicators of the disturbance response effect assessment mechanism include four items: compliance rate of fulfillment constraints, fresh produce loss rate, delivery timeliness achievement rate, and additional scheduling costs. After a Level 1 disturbance response, the core hard constraints must be 100% compliant. After a Level 2 disturbance response, the delivery timeliness achievement rate must be no less than 95%. After a Level 3 disturbance response, no additional scheduling costs are incurred. If the evaluation indicators do not meet the standards, a secondary adjustment strategy will be initiated immediately.

[0043] The risk warning stage is based on the real-time performance data, disturbance identification results and response effect evaluation data collected in the disturbance response stage. Combined with historical performance big data, fresh food category characteristic data and environmental prediction data, a performance default risk prediction model based on LSTM long short-term memory network and reinforcement learning is constructed. This model can achieve self-adaptive optimization to improve prediction accuracy. The performance default risk prediction model predicts the full-chain performance default probability of each order and each order delivery group in real time. The model input dimensions include various relevant indicators such as performance progress deviation, road traffic prediction data, cold chain equipment operation status, fresh food loss trend, and user feedback information. The LSTM network of the default risk prediction model has three hidden layers with 128, 64, and 32 neurons respectively. The activation function is ReLU, and the output layer uses the Sigmoid function to output the default probability value. Reinforcement learning uses each fulfillment node of fresh food delivery as an agent and a real-time fulfillment scenario as the environment. The reward function is R = 1 - real-time default probability. The model training uses batch gradient descent with a batch size of 32 and a learning rate of 0.001. Training stops when the loss function value is ≤0.01. The model evaluation is based on AUC ≥0.9 and prediction accuracy ≥85% as the passing standard. If the model evaluation is unsatisfactory, the number of neurons in the hidden layers of the LSTM network or the learning rate of reinforcement learning are immediately readjusted, and batch gradient descent is used for retraining until the model evaluation meets the passing standard.

[0044] The fulfillment rate refers to the proportion of orders that were delivered without breaking any hard constraints or exceeding risk constraints to the total number of delivered orders; the actual spoilage rate of fresh produce refers to the proportion of fresh produce orders that were spoiled during delivery to the total number of delivered orders, calculated as the ratio of the actual spoilage weight to the total order weight; the unit delivery cost rate refers to the ratio of the average delivery cost per order to the average delivery revenue per order. All three indicators are statistically analyzed separately for each delivery task, each product category, and each scenario, providing accurate numerical data for parameter optimization.

[0045] Set thresholds for general risk, higher risk, and major risk. When the predicted probability of default exceeds the corresponding risk threshold, a tiered early warning mechanism is triggered, and multiple risk handling plans are generated simultaneously. These plans include various methods such as capacity adjustment, route optimization, user communication, and compliance remediation. The multi-level risk thresholds are quantified according to the probability of default: 20%-40% is the general risk range, 40%-70% is the high risk range, ≥70% is the major risk range, and <20% is the risk-free range. Different ranges trigger corresponding levels of early warning signals and emergency response plan generation rules.

[0046] Based on the principles of minimizing disposal costs, maximizing performance assurance, and minimizing impact on user experience, a multi-objective decision-making algorithm is used to select the optimal disposal plan and execute it automatically, while the execution effect of the disposal plan is tracked in real time. If the probability of default remains higher than the threshold after the measures are taken, an upgraded handling process will be triggered, such as activating higher-priority standby capacity or platform intervention to coordinate and ensure that core performance constraints are not breached.

[0047] The escalation process for handling defaults begins by activating the platform's dedicated emergency transport capacity, utilizing high-rated emergency cold chain transport to undertake the delivery task. Next, a dedicated person communicates one-on-one with the user, synchronizing delivery adjustments and providing a reasonable compensation plan. Finally, the source of default risk is monitored in real time; if the issue is with cold chain equipment, on-site repairs are arranged; if it's a transportation problem, dedicated routes are coordinated. The entire escalation process is synchronized to the platform's dispatch center in real time until the probability of default drops to a safe range.

[0048] When selecting contingency plans using a multi-objective decision-making algorithm, appropriate weights are first assigned to three principles: handling cost, performance guarantee effect, and user experience impact. Then, each contingency plan is quantitatively scored across these three dimensions, and a comprehensive score is calculated by weighting the scores. The contingency plan with the highest comprehensive score is the optimal handling solution. If there are identical comprehensive scores, the contingency plan with the higher performance guarantee effect score is prioritized. The selection process is completed automatically by the system without any manual intervention, and the selection results are synchronized to the scheduling execution end in real time. The system automatically completes disturbance identification and triggers graded response strategies. Operations and scheduling personnel can view disturbance details. Manual intervention is only required to adjust the response strategy if the handling effect of a Level 1 major disturbance fails to meet the standards.

[0049] The platform's emergency delivery capacity is reserved in zones according to delivery areas. The number of emergency delivery capacity in each area is no less than 15% of the total daily delivery capacity in that area, and all of them are S / A-level high-rated cold chain delivery capacity, fully equipped with backup refrigeration equipment and emergency delivery supplies. Emergency delivery capacity is dispatched according to the priority of default risk level, delivery area, and order category. Orders with major risks are dispatched first, followed by orders of cold chain sensitive categories such as seafood and meat. Orders of the same level and category are sorted according to delivery time requirements. The dispatch results are synchronized to the emergency delivery capacity end and the order user end in real time.

[0050] In the iterative optimization phase, after each delivery task is completed, the closed-loop management results and handling data from the risk warning phase are combined to collect full-chain actual performance data, including the completion status of performance nodes, the execution status of constraints, the actual loss rate of fresh produce, the performance of transportation capacity, user satisfaction, the cost of handling defaults, and other relevant data, and to establish a performance data knowledge base. The performance data knowledge base is built on a three-level classification architecture based on category, scenario, and stage. The category dimension includes categories such as fruits and vegetables, meat, seafood, and pre-cooked meals. The scenario dimension includes categories such as community delivery, supermarket delivery, and instant delivery. The stage dimension includes the entire process stages such as constraint modeling, aggregation matching, and capacity matching. The knowledge base data supports multi-dimensional combination retrieval, providing targeted data support for model parameter optimization.

[0051] The core data sources of the performance data knowledge base cover four major categories: full-chain performance monitoring data, model algorithm operation data, disturbance response and risk management data, and user feedback and capacity performance evaluation data. All data sources are automatically connected. The knowledge base is updated incrementally according to the delivery task completion time. Relevant data is entered synchronously after each delivery task is completed. A full data verification and cleaning is performed monthly to remove invalid and erroneous data to ensure the integrity and accuracy of the knowledge base data. When each order is incrementally entered, the system automatically verifies the data integrity (such as whether performance node data and loss rate data are missing) and rationality (such as whether the loss rate exceeds the reasonable range of 0-100%). Data that fails the verification is marked as abnormal and returned to the entry point. A full data verification and cleaning is performed again monthly.

[0052] By comparing and analyzing actual performance data with preset performance constraints and optimization objectives, optimization points are identified in constraint models, order aggregation algorithms, capacity matching rules, route planning algorithms, and risk warning models. Performance differences for different fresh food categories and delivery scenarios (such as community delivery, supermarket delivery, and instant delivery) are also identified. Through reinforcement learning algorithms, various relevant parameters, such as thresholds and weights of performance constraint models, parameters of order aggregation algorithms, weight allocation of capacity matching, target weights of route planning, and thresholds of risk warnings, are adaptively and iteratively optimized to achieve dynamic adjustment of parameters. The reinforcement learning algorithm adopts the Q-Learning algorithm framework, which is a different specific application of the same technical system as the reinforcement learning in the risk warning stage. Both take minimizing the risk of breach of contract as the core optimization objective. The state space is the reasonable value range of each model parameter, the action space is the parameter adjustment step size, and the reward function is R = fulfillment qualification rate - actual fresh produce loss rate - unit delivery cost rate. The parameter iteration update frequency is once every 100 delivery tasks completed. After the parameter update, it is verified by the actual fulfillment data of the subsequent 30 new delivery tasks. If the fulfillment qualification rate decreases by ≥5% after optimization, the parameter is confirmed to be effective; otherwise, it is rolled back to the original parameter.

[0053] The optimized model parameters are synchronously updated to the digital twin model of performance constraints in the constraint modeling stage, the order aggregation algorithm in the aggregation matching stage, the path planning algorithm in the path planning stage, and the early warning model in the risk warning stage; a scenario adaptation module is established to automatically call the optimized parameters and models for different fresh food categories and different delivery scenarios, adapt to the dynamic changes in fresh food delivery scenarios, and continuously improve the fulfillment rate, reduce operating costs and fresh food spoilage.

[0054] The scene adaptation module has a built-in category and scene feature recognition engine, which can automatically extract the fresh food category attributes and delivery scene attributes of the delivery task, and build a dedicated parameter model library according to the combination of fruit and vegetable + instant delivery, seafood + supermarket delivery, etc. After recognition, it automatically matches the optimized parameters and models of the corresponding combination. At the same time, it supports manual fine-tuning of parameters and models for special categories and special scenes, and the fine-tuning results are updated to the model library synchronously.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent scheduling of fresh food delivery based on fulfillment constraints, characterized in that, The specific steps include the following: In the constraint modeling phase, a multi-dimensional dynamic performance constraint indicator system covering four dimensions—category characteristics, user needs, compliance requirements, and capacity limits—is constructed. A three-level constraint control system is defined, and the cost of breach of contract and trigger thresholds are quantified. A digital twin model of performance constraints and a constraint conflict prediction mechanism are also constructed. During the aggregation and matching phase, the pool of fresh food orders to be delivered is obtained and relevant parameters of order fulfillment constraints are extracted. User fulfillment preferences and compliance priorities are incorporated, and orders are dynamically aggregated and grouped through the order aggregation algorithm model. Control standards are set, and an order aggregation taboo and dynamic adjustment mechanism is established. During the capacity matching phase, a digital profile system and a capacity rating system based on fulfillment capability are constructed. Two-level matching verification is performed on order delivery groups and capacity, and the comprehensive matching degree is calculated to establish a dynamic capacity adaptation mechanism. In the path planning phase, a performance constraint game optimization function is constructed with minimizing the risk of breach of performance as the core and multiple auxiliary objectives working together for optimization. A multi-objective optimization algorithm that adapts to constraint priority rules is used to solve for the optimal solution, and the management requirements for performance nodes are clarified. During the disturbance response phase, data from multiple systems are integrated, status data of all performance nodes across the entire chain are collected, performance disturbance events are identified and disturbance levels are classified, corresponding response strategies are triggered for different levels of disturbances, and a disturbance response effectiveness evaluation mechanism is established. During the risk warning phase, a performance default risk prediction model is constructed by combining various relevant data, multi-level risk thresholds are set and tiered warnings are triggered, multiple risk handling plans are generated, and the optimal solution is selected based on the multi-objective decision-making principle to establish a closed-loop management and control mechanism for the entire process. In the iterative optimization phase, after the delivery task is completed, actual fulfillment data of the entire chain is collected and a fulfillment data knowledge base is established. Optimization points of models and algorithms at each stage are explored, and reinforcement learning algorithms are used to achieve adaptive iterative optimization of the parameters of each model and update the relevant models at each stage.

2. The intelligent scheduling method for fresh food delivery based on fulfillment constraints according to claim 1, characterized in that, In the constraint modeling phase, the core fulfillment nodes of the entire fresh food delivery chain, from order generation to user signature, are identified. A fulfillment constraint indicator library is built for each node, and each indicator is linked to fresh food category lifecycle parameters and real-time environmental data to achieve dynamic updates. The three-level constraint control system is based on the inviolability of constraints and the degree of impact of breach. Hard constraints include inviolable requirements such as mandatory temperature zones for fresh food categories, food safety compliance, and complete traceability information. Soft constraints include content such as user-adjustable signature time and delivery order preferences. Risk constraints use thresholds such as fresh food spoilage rate and delivery capacity fatigue as the basis for pre-control. The fulfillment constraint digital twin model integrates real-time data perception functions to achieve digital control and dynamic updates of constraints throughout the entire chain. The constraint conflict prediction mechanism is used to identify potential conflicts between constraints at different fulfillment nodes in advance.

3. The intelligent scheduling method for fresh food delivery based on fulfillment constraints according to claim 2, characterized in that, In the aggregation and matching phase, based on the constraint indicator library, hierarchical control system, and conflict prediction results from the constraint modeling phase, the performance constraint parameters, constraint levels, and priority weights of each order are extracted. These are then integrated with user performance preference profiles and compliance priorities. A weighted K-means clustering algorithm model, which combines constraint compatibility verification factors and default risk prediction functions, is used to achieve dynamic order aggregation and grouping. The order aggregation taboo rules are used to prohibit order aggregation operations where the probability of default after merging exceeds a preset threshold, violates compliance priorities, or conflicts with user preferences. The compliance priority prioritizes food safety compliance, with other compliance dimensions dynamically adjusted based on the characteristics of fresh produce categories. For cold chain sensitive categories, the priority of cold chain temperature control compliance is increased. The dynamic adjustment mechanism is used to receive information including new orders, order cancellations, and changes in user requirements, and to split and reorganize the aggregated order group.

4. The intelligent scheduling method for fresh food delivery based on fulfillment constraints according to claim 3, characterized in that, In the capacity matching phase, a digital profile system for the cold chain delivery capacity resource pool is constructed, comprising three dimensions: physical attributes, historical fulfillment attributes, and real-time status. The capacity fulfillment capability rating system is a four-level rating system, with the rating results dynamically correlated with the data in the three dimensions: physical attributes, historical fulfillment data, and real-time status. In the two-level matching verification, the first level is a hard constraint veto verification, filtering out capacity that does not meet the core hard constraints of the order group. The second level is a weighted calculation of the comprehensive matching degree between the remaining capacity and the order delivery group. The dynamic capacity adaptation mechanism is used to monitor the capacity operation status in real time, and to initiate the backup capacity matching process when a capacity operation anomaly occurs.

5. The intelligent scheduling method for fresh food delivery based on fulfillment constraints according to claim 4, characterized in that, In the path planning phase, the multi-objective performance constraint game optimization function takes minimizing the risk of performance default as its core objective, while also considering four auxiliary objectives: minimizing total delivery cost, minimizing fresh produce spoilage, optimizing delivery timeliness, and optimizing user experience, and clearly defining the weight of each objective. The constraints of the optimization function include full-link node constraints, order group constraints, and capacity constraints. Hard constraints serve as insurmountable boundary conditions, while soft constraints and risk constraints serve as dynamic adjustment factors. The multi-objective optimization algorithm integrates game theory to construct a three-party game model involving the platform, capacity, and users. After balancing the interests of the three parties, it solves for the Pareto optimal solution set based on the priority of order group performance constraints, capacity rating, and user preferences, and selects the globally optimal delivery route scheme. At the same time, it clarifies the control requirements for each performance node and dynamically adjusts the path and node thresholds by combining real-time traffic prediction models and fresh produce spoilage prediction models.

6. The intelligent scheduling method for fresh food delivery based on fulfillment constraints according to claim 5, characterized in that, During the disturbance response phase, status data of all performance nodes across the entire chain is collected through multi-system data acquisition and compared with preset performance node control thresholds to identify disturbance events. These performance disturbance events are classified into three levels based on their impact on performance constraints, scope, and duration: Level 1 is a major disturbance where hard constraints are about to be breached; Level 2 is a moderate disturbance where soft constraints exceed preset thresholds; and Level 3 is a minor disturbance that does not affect core performance. Response strategies, including emergency handling, dynamic adjustment, and early warning, are triggered for different levels of disturbances to control additional costs without breaching core performance constraints and balancing the interests of all parties. A disturbance response effect evaluation mechanism is used to track the adjustment effect in real time.

7. The intelligent scheduling method for fresh food delivery based on fulfillment constraints according to claim 6, characterized in that, During the risk warning phase, a performance default risk prediction model is constructed by combining real-time performance data, historical performance big data, fresh produce category characteristic data, and environmental prediction data. This model is used to predict the probability of performance default across the entire chain of orders and order delivery groups in real time. The multi-level risk threshold triggers a tiered warning mechanism, which simultaneously generates multiple risk handling plans, including capacity adjustments, route optimization, and user communication. Based on the principles of lowest handling cost, optimal performance guarantee effect, and minimal impact on user experience, the optimal plan is selected and executed through a multi-objective decision-making algorithm. The handling effect is tracked in real time. If the default probability still exceeds the preset risk threshold, an escalation handling process is triggered. The escalation handling process includes core steps such as high-priority backup capacity deployment, platform intervention and user communication, and real-time control of the source of default risk.

8. The intelligent scheduling method for fresh food delivery based on fulfillment constraints according to claim 7, characterized in that, During the iterative optimization phase, the performance data knowledge base collects actual performance data and closed-loop management and disposal data from the entire chain. By comparing the actual performance data with preset standards, it identifies optimization points for each model and algorithm and identifies performance differences for different categories and scenarios. Through reinforcement learning algorithms, it achieves adaptive iterative optimization of the parameters of each model, updates the optimized parameters to the relevant models at each stage, and establishes a scenario adaptation module to automatically call the optimized parameters and models for different fresh food categories and delivery scenarios.

9. The intelligent scheduling method for fresh food delivery based on fulfillment constraints according to claim 8, characterized in that, The user fulfillment preference profile is built based on the user's historical delivery order data over the past 90 days and is dynamically updated on a monthly basis.

10. The intelligent scheduling method for fresh food delivery based on fulfillment constraints according to claim 9, characterized in that, The fulfillment data knowledge base is built using a three-level classification architecture of category, scenario, and stage, and is updated incrementally according to the delivery task completion time. Full data verification and cleaning are carried out every month to remove invalid and erroneous data.

Images