Neural combination optimization model-fused perishable food distribution system and method

Abstract

The invention discloses a perishable food distribution system and method fused with a neural combination optimization model. The perishable food distribution system comprises a data simulation preprocessing unit, a virtual customer division unit, a maximum group grouping unit, a neural combination path planning unit, a complete path merging unit and a large neighborhood search optimization unit. The method comprises the following steps: firstly, through data preprocessing and virtual customer splitting, converting a multi-article demand customer into a single-demand virtual customer; taking temperature compatibility as a constraint, and grouping by using a maximum clique algorithm; an initial path is planned for each group through a pre-trained neural combination optimization model, and optimization is carried out through a large neighborhood search algorithm of an improved operator after splicing. According to the method, efficient search of neural combinatorial optimization and accurate optimization capacity of large neighborhood search are fused, delivery with temperature compliance, optimal path and lowest cost of perishable food is realized, multi-scene requirements are met, and delivery efficiency and quality are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of perishable food logistics distribution optimization and artificial intelligence, in particular to a perishable food distribution system and method fusing a neural combination optimization model. BACKGROUND

[0002] With the increasing demand for fresh, dairy and other perishable food by residents, the cold chain distribution industry continues to expand, but the existing distribution scheme faces significant technical bottlenecks: Multi-temperature requirement adaptation difficulty: A single customer often needs multiple perishable foods with different temperature requirements (such as refrigerated meat, low-temperature dairy products, and room-temperature cooked food) at the same time. Traditional solutions are difficult to balance different temperature compliance, which easily leads to food spoilage; Low path planning quality: Existing heuristic algorithms (such as the saving algorithm and the ant colony algorithm) are prone to local optimal solutions when dealing with multiple constraint scenarios, resulting in path redundancy, low vehicle utilization, and high distribution costs; High dependence on initial solution: Traditional algorithms highly depend on the quality of the initial solution. Without a high-quality initial solution, a large amount of computing resources is needed to obtain a usable solution, making it difficult to meet real-time distribution requirements; Insufficient adaptation of intelligent algorithms: Existing intelligent optimization methods are mostly suitable for simple vehicle path planning and cannot directly address the complex requirements of perishable food temperature constraints and quality loss.

[0003] In summary, how to achieve efficient and low-cost distribution of perishable food while meeting multi-temperature compliance has become a key problem in the cold chain logistics field that needs to be solved. SUMMARY

[0004] To solve the above problems, the present application aims to propose a perishable food distribution system and method fusing a neural combination optimization model, which can achieve efficient, accurate, and temperature-controlled distribution of perishable food by combining the efficient search capability of the neural combination optimization method and the precise optimization capability of the large neighborhood search algorithm.

[0005] To achieve the above purpose, the technical solution of the present application is as follows: A perishable food distribution system fusing a neural combination optimization model, comprising a data simulation preprocessing unit, a virtual customer division unit, a maximum group grouping processing unit, a neural combination path planning unit, a complete path merging unit, and a large neighborhood search optimization unit, which are connected in sequence; The data simulation preprocessing unit is used to collect and process customer information, perishable food demand information, and food physical property parameters to form structured basic data; The virtual customer division unit is used to split actual customers with multiple item demands into virtual customers with single item demands; The maximum group grouping processing unit is configured to group the virtual customers according to temperature compatibility; The neural combination path planning unit is configured to plan an initial path for each group of virtual customers by using a pre-trained neural combination optimization model; The complete path merging unit is configured to splice the initial paths of each group to form a complete initial delivery scheme; The large neighborhood search optimization unit is configured to optimize the complete initial delivery scheme by using an improved operator.

[0006] Further, the data processed by the data simulation preprocessing unit includes customer geographic location coordinates (x, y) , , delivery time range (t1, t2) , , item type and demand, minimum tolerance temperature of food transportation , maximum tolerance temperature , closed-door transportation quality loss rate ɑ, open-door transportation quality loss rate ; the processing process includes abnormal data cleaning and missing data interpolation.

[0007] Further, each virtual customer split by the virtual customer division unit retains the original actual customer's geographic location coordinates, delivery time range, and temperature requirement range of the corresponding item (x, y, t1, t2) , .

[0008] Further, the grouping constraint of the maximum group grouping processing unit is that the intersection of the temperature requirement ranges of all virtual customers in the same group is non-empty, and the grouping process includes constructing a temperature compatibility graph, initializing an ungrouped set and a grouped set, and iteratively extracting a maximum group and updating the sets.

[0009] Further, the neural combination optimization model of the neural combination path planning unit adopts an encoding-decoding structure, the encoding module is a graph neural network (GNN), and the decoding module is a Transformer structure with an attention mechanism; the training target is to minimize the total mileage, the number of refrigerated vehicles used, and the food quality loss; and the hyperparameters can be automatically adjusted according to the group size.

[0010] Further, the splicing method of the complete path merging unit is to use the first group path as the base segment, and the start point of each subsequent group path is connected to the end point of the previous group path without conflict coordination.

[0011] Further, the large neighborhood search optimization unit takes the complete initial delivery scheme as the optimization object, cooperatively operates the improved destruction operator and the improved repair operator, and realizes deep optimization of the path by using the simulated annealing acceptance criterion.

[0012] In order to achieve the above object, the application further provides a perishable food distribution method based on a neural combination optimization model, comprising the following steps: S1: data simulation preprocessing: collecting and processing customer information, perishable food demand information and food physical attribute parameters to form structured data; S2: virtual customer division: splitting the actual customers with multi-item demand into virtual customers with single-item demand; S3: maximum group grouping: grouping the virtual customers according to temperature compatibility; S4: neural combination path planning: planning an initial path for each group by using a pre-trained neural combination optimization model; S5: complete path merging: splicing the initial paths of each group to form a complete initial scheme; S6: large neighborhood search optimization: optimizing the complete initial scheme by using an improved operator to output an optimal distribution path.

[0013] Further, the temperature compatibility graph construction method of the maximum group grouping in step S3 is as follows: initializing an undirected graph G=(V,E), V is a virtual customer set, if the temperature range intersection of virtual customers i and j is not empty, then an edge (i,j) is added between i and j to the edge set E.

[0014] Further, the improved operator optimization in step S6 includes calculating a benchmark cost, and the benchmark cost calculation formula is Cost= × total mileage of the path + × number of refrigerated vehicles used + × mass loss during transportation, wherein 、 and are weight coefficients; the comprehensive cost: the simulated annealing acceptance probability formula is P= , T is a temperature parameter of the current iteration number (T= × , wherein k is the current iteration round, k∈[1,200].

[0015] Beneficial effects: (1) Efficient and accurate delivery path planning and temperature compliance guarantee: By combining the global efficient initial solution generation capability of the pre-trained neural optimization model and the local precise optimization capability of the improved large neighborhood search algorithm, the present application can realize the whole-link delivery path planning of perishable food under multiple temperature constraints. Compared with traditional single heuristic algorithm, it can quickly find the optimal path considering cost and compliance in the complex scene of "one customer has multiple item demands, and different items require different transportation temperatures", significantly improving planning efficiency; at the same time, through the temperature compatibility constraint of the maximum group grouping and the temperature verification mechanism after path generation, it ensures that the items are within the required temperature range throughout the whole process, the temperature compliance rate remains 100%, effectively avoiding the deterioration of items caused by ignoring temperature constraints in traditional methods, and guaranteeing the delivery quality and safety.

[0016] (2) Fine-grained demand processing and efficient use of delivery resources: When building the delivery model, the temperature demand range, geographic location coordinates, delivery time window, and item demand of virtual customers are carefully disassembled and accurately represented, and multi-feature fusion is realized through graph neural network (GNN). Taking "the shortest total mileage + the least number of refrigerated vehicles used + the least quality loss during transportation" as the multi-objective function, the delivery cost minimization and vehicle utilization maximization are realized, reducing the resource waste caused by rough demand processing in traditional methods, fully utilizing the transportation efficiency of cold chain vehicles, and reducing the comprehensive operating cost of perishable food delivery.

[0017] (3) Intelligent path optimization adjustment and quality loss minimization guarantee: The local path optimization using the improved large neighborhood search algorithm is highly targeted, accurately locating the inefficient nodes in the path through destruction operators (such as similarity-based removal, maximum quality loss reduction removal), and adjusting the path order within the constraint range through repair operators (such as minimum quality loss insertion, time window relief insertion).

[0018] (4) Wide scene applicability and good scalability: The present application fully considers different scales and types of perishable food delivery scenarios in design, whether it is small community fresh food delivery (30-50 customers) or large city multi-region delivery (150-200 customers), the model adaptation module can automatically adjust the hyperparameters to ensure efficient landing of the solution; at the same time, with the upgrading of delivery demand (such as the addition of frozen item temperature demand, the introduction of unmanned aerial vehicle assisted delivery), the virtual customer feature dimension can be conveniently expanded (such as adding item freezing temperature threshold), and the optimization algorithm operator can be conveniently expanded (such as adding unmanned aerial vehicle path adaptation repair operator), without the need to restructure the system framework, adapting to the technical development trend of the perishable food delivery industry, and ensuring the long-term effectiveness of the solution.

[0019] (5) Wide scene applicability and good expansibility: the application fully considers different scales and types of perishable food distribution scenes in the design, whether it is small community fresh food distribution (10-20 virtual customers) or large city multi-region distribution (30-50 virtual customers), the model adaptation module can automatically adjust the hyperparameters to ensure that the scheme is efficiently landed; at the same time, as the distribution demand upgrades (such as adding frozen food temperature requirements, introducing unmanned aerial vehicle auxiliary distribution), the virtual customer feature dimension can be conveniently extended (such as adding frozen food temperature thresholds), and the algorithm operator can be optimized (such as adding unmanned aerial vehicle path adaptation repair operator), without the need to reconstruct the system framework, adapt to the technical development trend of the perishable food distribution industry, and ensure the long-term effectiveness of the scheme.

[0020] In summary, through rigorous virtual customer data processing, accurate multi-objective distribution model construction, efficient fusion algorithm optimization and comprehensive quality verification, the application ensures that the perishable food distribution process realizes "temperature compliance, cost optimization and highest efficiency", effectively deals with multi-constraint and multi-variable distribution challenges, improves the overall operation performance of the perishable food distribution system, and provides strong support for the high-quality development of the cold chain logistics industry. BRIEF DESCRIPTION OF DRAWINGS

[0021] The accompanying drawings, which form a part of this application, are used to provide a further understanding of the application, and the illustrative embodiments thereof, and are not intended to limit the application. In the drawings: Figure 1 A functional structure diagram of a perishable food distribution system of a fusion neural combination optimization model according to an embodiment of the application is shown in the figure. Figure 2 A flowchart of a perishable food distribution method of a fusion neural combination optimization model according to an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0022] It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0023] The application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0024] Embodiment 1 Referring to Figure 1 A perishable food distribution system of a fusion neural combination optimization model, comprising a data simulation preprocessing unit, a virtual customer division unit, a maximum group grouping processing unit, a neural combination path planning unit, a complete path merging unit and a large neighborhood search optimization unit, each unit is connected in turn. The data simulation preprocessing unit is used to collect and process customer information, perishable food demand information and food physical attribute parameters to form structured basic data. The virtual customer dividing unit is configured to divide actual customers with multi-item demands into virtual customers with single-item demands; The maximum group grouping processing unit is configured to group the virtual customers according to temperature compatibility; The neural combination path planning unit is configured to plan initial paths for each group of virtual customers by using a pre-trained neural combination optimization model; The complete path merging unit is configured to splice the initial paths of each group to form a complete initial delivery scheme; The large neighborhood search optimization unit is configured to optimize the complete initial delivery scheme by using an improvement operator.

[0025] The embodiment combines the global efficient initial solution generation capability of the pre-trained neural combination optimization model and the local precise optimization capability of the improved large neighborhood search algorithm, and realizes the full efficient delivery path planning of perishable food with multiple constraints. Compared with traditional single heuristic algorithms, the embodiment can quickly find a delivery scheme with the highest customer satisfaction and the lowest cost in a complex scenario where one customer has multiple item demands and different items require different transportation temperatures, and significantly improves the planning efficiency and comprehensive quality of perishable food delivery.

[0026] In a specific example, the data simulation preprocessing unit processes data including customer geographic location coordinates (x i, y i) , , delivery time range (t i, T i) , , item types and demand, minimum tolerable temperature for food transportation , maximum tolerable temperature , closed-door transportation quality loss rate ɑ, open-door transportation quality loss rate . The processing process includes abnormal data cleaning and missing data interpolation.

[0027] It should be noted that by conducting multi-dimensional research on the market demand of the target delivery area, customer information (including the specific geographic location coordinates (x i, y i) , of customer i, the delivery time range (t i, T i) , required by customer i, the service time of customer i , market demand data of various perishable foods (covering the type details of items required by customers, the specific demand of each item), and the physical attribute parameters of different types of perishable foods (including the minimum tolerable temperature , the maximum tolerable temperature , the quality loss rate ɑ during closed-door transportation of the refrigerated truck, and the quality loss rate during open-door transportation of the refrigerated truck) are collected and integrated.); based on the research data, a simulation data set is constructed, abnormal data (such as temperature requirements exceeding a reasonable range, incorrect geographic location information) is cleaned and verified, missing data is supplemented by using an average value, and a structured and standardized basic data asset is formed, thereby providing accurate, complete and usable data support for subsequent virtual customer division, grouping processing and path planning.

[0028] In a specific example, the virtual customer division unit splits each virtual customer to retain the geographic location coordinates, delivery time range, and temperature requirement range of the corresponding item of the original actual customer , ).

[0029] It should be noted that for the multi-item demand of each actual customer, the split processing is performed according to the item category, and a single actual customer with multiple item demands is converted into multiple virtual customers, each of which is associated with only one item demand, and retains the geographic location information of the original actual customer and the transportation temperature requirement range of the item, thereby achieving fine disassembly of customer demand and providing a basic unit for subsequent grouping processing based on temperature compatibility. For example, for customer A, the location coordinates of the customer are (30, 40), the service time is required to be (7:00, 12:00), and two items, lemons and frozen fruits, are required, wherein the demand for lemons is 20 kg, the minimum storage temperature during transportation is required to be 10 degrees, and the maximum storage temperature is required to be 15 degrees, the demand for frozen fruits is 42 kg, and the minimum storage temperature during transportation is required to be -30 degrees, and the maximum storage temperature is required to be -15 degrees. This customer is split into two virtual customers and , which have the same location coordinates and service time as customer A, virtual customer requires 20 kg of lemons, and virtual customer requires 42 kg of frozen fruits.

[0030] In a specific example, the grouping constraint of the maximum clique grouping processing unit is that the intersection of the item temperature requirement ranges of all virtual customers in the same group is non-empty, and the grouping process includes constructing a temperature compatibility graph, initializing an ungrouped set and a grouped set, and iteratively extracting a maximum clique and updating the sets.

[0031] It should be noted that the virtual customer division unit outputs all virtual customer data, each of which corresponds to a single item category and carries the transportation temperature requirement range of the item , (j is the virtual customer identifier). Then, grouping is performed using the maximum clique algorithm according to the following steps, and always taking the constraint that "there is a non-empty intersection between the temperature range of all virtual customer demand items in the same group" as the core, to ensure that the grouping result meets the temperature control compatibility requirements in subsequent single-vehicle transportation: Temperature compatibility graph construction: initialize the undirected graph G = (V, E), where the vertex set V = {1, 2,..., n} (n is the total number of virtual customers, and the vertex represents the virtual customer index). Traverse all virtual customer pairs (i, j) (i = 1 to n, j = i + 1 to n), if the intersection of the temperature range of virtual customer i [ , ] and the temperature range of virtual customer j [ , ] is non-empty, i.e. , ∩ , ≠ , then add an edge (i, j) between vertex i and j to the edge set E, to complete the construction of the temperature compatibility graph. For example, for four virtual customers with temperature ranges [7, 14], [9, 23], [-12, -1], and [-5, 5], the constructed temperature compatibility graph is: .

[0032] Initialization of ungrouped and grouped sets: initialize the ungrouped virtual customer set: U ← V (at this time V = {1, 2, 3, 4}, corresponding to the above four virtual customers), and the grouped set Groups ← . When U ≠ , perform the following operations: extract induced subgraph: extract the subgraph induced by the vertices in the current U from the graph G . Initially, U = {1, 2, 3, 4}, so is the complete temperature compatibility graph corresponding to the subgraph.

[0033] Iterative extraction of maximum cliques and updating of sets: find maximum cliques: find maximum cliques in the subgraph (the maximum clique is the largest subset of vertices in the subgraph that are connected to each other). Observing the adjacency matrix of the above temperature compatibility graph, there is an edge between vertex 1 and vertex 2 (the elements in the 1st row and 2nd column, and 2nd row and 1st column of the matrix are 1), and there is an edge between vertex 3 and vertex 4 (the elements in the 3rd row and 4th column, and 4th row and 3rd column of the matrix are 1), but there is no edge between vertices 1, 2 and vertices 3, 4 (the corresponding matrix elements are 0). Therefore, there are two maximum cliques, which are = {1, 2} and ={3,4} (there are edges between every two vertices in these two subsets, and no other vertices can be added to form a larger subset of two-by-two connected vertices); update the grouped and ungrouped sets: randomly select one of the largest cliques (for example, select the first one ={1,2}), add it to the grouped set, that is, Groups←{Groups}∪ }; then remove the grouped virtual customers from U, at this time U←{3,4}. Execute the above steps again, extract the subgraph induced by U={3,4} , find the largest clique in the subgraph, obtain ={3,4}, add it to the grouped set Groups←{Groups}∪ , and update U← .

[0034] In a specific example, the neural composition path planning unit adopts a neural composition optimization model with an encoding-decoding structure, the encoding module is a graph neural network GNN, and the decoding module is a Transformer structure with an attention mechanism; the training target is the shortest total mileage, the least number of refrigerated vehicles used, and the least loss of food quality; and the hyperparameters can be automatically adjusted according to the group size.

[0035] It should be noted that the unit takes the "same-temperature intersection virtual customer group" obtained after maximum group grouping as the input object, and uses a pre-trained neural composition optimization model to realize path planning. Since the required transportation temperature ranges of all virtual customers in the same group have intersections, we eliminate the temperature constraint in cold chain transportation in the path planning process within the group, and we can regard it as a normal vehicle path planning problem. The neural composition optimization model adopts an encoding-decoding structure, the encoding module fuses and represents the geographical location coordinates, delivery time window constraints, the amount of each virtual customer's demand, and the total capacity of the refrigerated vehicle through a graph neural network (GNN), and converts the customer delivery demand into a high-dimensional feature vector; the decoding module adopts a Transformer structure with an attention mechanism, dynamically generates a path sequence through a self-recurrent manner, and takes "the shortest total mileage + the least number of refrigerated vehicles used + the least loss of quality during transportation" as a multi-objective optimization function during the generation process, and calculates the comprehensive cost of the current path in real time and feeds back to adjust the generation strategy. At the same time, the unit also has a built-in model adaptation module, which can automatically adjust the model hyperparameters (such as the number of attention heads and the hidden layer dimension) according to the size of the virtual customer group (such as 10-50 virtual customers), ensuring that accurate paths are quickly generated for small-scale groups, and efficiency and solution quality are considered for large-scale groups, and finally outputting an initial path for each virtual customer group that meets the temperature constraint and basic delivery requirements.

[0036] In a specific example, the complete path merging unit is spliced ​​as follows: the first group of paths is used as the base segment, and the starting point of each subsequent group of paths is connected to the ending point of the previous group of paths, without any conflict coordination.

[0037] It should be noted that this unit is used to directly concatenate the complete paths of each virtual customer group output by the neural combined path planning unit (0 represents a warehouse, and the remaining numbers represent different virtual customers). Since the grouped paths output by the neural combined path planning unit have pre-satisfied geographical connectivity, temperature zone compatibility, and vehicle load adaptability, there are no conflicts before concatenation, so no conflict coordination operation is required. The specific process is as follows: Assuming the neural combined path planning unit outputs three grouped paths as [0,2,7,3,0,5,6,9,0], [0,11,1,4,0,2,8,10,0], and [0,15,13,0,12,16,0], this unit first concatenates the neural combined paths... The path planning unit outputs the grouped path sequence. The first grouped path [0,2,7,3,0,5,6,9,0] is used as the basic segment of the complete path. Then, the starting point (warehouse 0) of the second grouped path is directly connected to the ending point (warehouse 0) of the first grouped path to form [0,2,7,3,0,5,6,9,0,11,1,4,0,2,8,10,0]. Then, the starting point (warehouse 0) of the third grouped path is connected to the ending point (warehouse 0) of the previous connected path to form the preliminary complete path [0,2,7,3,0,5,6,9,0,11,1,4,0,2,8,10,0,15,13,0,12,16,0].

[0038] In a specific example, the large neighborhood search optimization unit takes the complete initial delivery plan as the optimization object, and achieves path depth optimization by improving the cooperative operation of the destruction operator and the repair operator, combined with the simulated annealing acceptance criterion.

[0039] It should be noted that the initial solution input and cost calculation are as follows: Input the initial complete path (including the item temperature range, geographical coordinates (x, y), delivery time window, and item quality loss coefficient for each virtual customer) to calculate the baseline cost. The formula is: Cost = ×Total route mileage+ ×Number of refrigerated vehicles used+ × Quality loss during transportation (of which) , and (Weighting coefficients, set according to specific circumstances), and this cost is used as the baseline solution for optimization. The total mileage is calculated using the Euclidean function; for example, for two virtual customers with coordinates (2, 5) and (3, 5), the distance between them is... , the sum of the distances between all consecutive customers is the total distance of the path. The number of refrigerated vehicles used is the number of warehouses 0 in the complete path minus 1, and for the complete path [0, 2, 7, 3, 0, 5, 6, 9, 0, 11, 1, 4, 0, 2, 8, 10, 0, 15, 13, 0, 12, 16, 0] the number of vehicles used is 6. The quality loss of the virtual customers during transportation is calculated by the formula = ·(1- ), where is the total time during which the required item of the virtual customer was in the refrigerated vehicle during the time the vehicle door was open, is the total time during which the required item of the virtual customer was in the refrigerated vehicle during the time the vehicle door was closed, and is the quality loss rate during transportation when the vehicle door is closed, and is the quality loss rate during transportation when the vehicle door is open. The quality loss of all virtual customers is added to obtain the quality loss during transportation.

[0040] Improved destruction operator operation: randomly select one destruction operator each iteration, remove some customer delivery nodes to form an incomplete path to be repaired, and the specific operations of each operator are as follows: Random removal operator: randomly remove 20% of the total number of virtual customers from the path, without bias to break the current path structure. Introduce randomness in the early iterations to avoid premature convergence of the algorithm to a local optimum and ensure the ability to explore the solution space. Similarity-based removal operator: first calculate the similarity between any two customer nodes and , the formula is: sim( , ) = a · simtem( , ) + (1 a) · simspa( , ) (where a ∈ [0, 1] is the balance coefficient, simtem is the temperature overlap similarity, and simspa is the spatial similarity). Temperature overlap similarity: simtem( , ) = (where , , and represent the minimum transportation temperature requirement, the maximum temperature requirement, the virtual customer Minimum and maximum transport temperature requirements for the required items. , The value range is [0,1], representing the proportion of the overlap length to the total span of the temperature range. (For virtual customers...) Minimum transport temperature requirements =0, maximum temperature requirement =15, Virtual Customers Minimum transport temperature requirements =3, Maximum temperature requirement =12, then calculate simtem according to the formula ( , )= = = ; Spatial similarity: simspa ( , )= (in( , )and( , These are virtual customers. and The closer the location coordinates, the closer the similarity (1). For virtual customers The position coordinates are ( , ) = (30, 55), virtual customers The position coordinates are ( , ) = (60, 15), then according to the formula, simspa( , )= = ; When the balance coefficient α is 0.2, the virtual customer and The similarity is sim( , ) = α·simtem ( , )+(1 α)·simspa( , )=0.2· +0.8· =0.13.

[0041] Customers are clustered based on the similarity matrix (within-group similarity being significantly higher than between-group similarity), and one cluster is randomly selected and all customer nodes within the group are removed to break the cluster substructure in the path; Random sub-path removal operator: identify the continuous sub-paths in the path corresponding to a single refrigerated vehicle service (e.g., [0, 2, 7, 3, 0] is one sub-path), randomly remove one complete sub-path, reduce redundant vehicle allocation, create space for subsequent re-allocation of removed customers to remaining vehicles, and improve overall fleet efficiency; Distance-optimal removal operator: evaluate the reduction in path mileage after removing each customer node one by one, preferentially remove the node that contributes most to the reduction in total mileage, and improve path efficiency through the "individual evaluation-preferential removal" framework; Quality loss maximum reduction removal operator: evaluate the reduction in total quality loss after removing each customer node, preferentially remove the node that contributes most to the reduction in quality loss, and ensure that the removal operation directly improves the quality performance of the delivery path.

[0042] Improved repair operator operation: for the incomplete path formed by the destruction operator, randomly select one repair operator to re-insert the removed node into the path, and strictly follow the vehicle temperature partition capacity, loading capacity, and time window constraints during the insertion process. The specific operations of each operator are as follows: Random feasibility constraint insertion operator: randomly select a position to insert the removed node in the feasible positions that satisfy the vehicle capacity constraints, temperature constraints, and time window constraints, enhance the diversity of reconstructed solutions, and avoid falling into local optima; Time window relief insertion operator: calculate the increment of the total waiting time of the path after inserting the removed node into each feasible position, preferentially select the position with the smallest increment for insertion, reduce the risk of time window conflicts, and ensure the punctuality of delivery; q-regret insertion operator: calculate the cost difference of the top q optimal positions before node insertion, consider the short-term cost increment and long-term regret risk, select the position with "small cost increment and low regret value" for insertion, and balance the solution quality and stability; Distance minimization insertion operator: calculate the path mileage increment after inserting the node into each feasible position, preferentially select the position with the smallest increment for insertion, and minimize the total travel distance under the premise of satisfying the constraints; Minimum quality loss insertion operator: calculate the quality loss increment after inserting the node into each feasible position, preferentially select the position with the smallest increment for insertion, and reduce the damage to the quality of the goods caused by path adjustment. Calculate the comprehensive cost of the new path: after repairing the incomplete path using the repair operator, a new path is obtained. If the comprehensive cost of the new path is lower than the benchmark solution, replace the benchmark solution directly; if the comprehensive cost of the new path is higher than the benchmark solution, use the simulated annealing acceptance criterion to decide whether to accept, and avoid falling into local optima. The acceptance probability decreases with the increase of the number of iterations, and the formula is P= , T is the temperature parameter of the current iteration number (T= × , where k is the current iteration round, k e [1, 200]; and l is a temperature attenuation coefficient, in the embodiment, l = 0.95, and the value range is [0.85, 0.99], which ensures that the temperature gradually decreases with the increase of the iteration number). If the new solution is accepted, the original benchmark solution is replaced with the new solution to continue the optimization until the termination condition is met (satisfying the maximum iteration number 200 times).

[0043] Embodiment 2 To achieve the above purpose, see Figure 2 : The present embodiment also provides a perishable food distribution method based on the neural combination optimization model, comprising the following steps: S1: Data simulation preprocessing: collecting and processing customer information, perishable food demand information and food physical attribute parameters to form structured data; S2: Virtual customer division: splitting the actual customer with multiple item demands into virtual customers with single item demands; S3: Maximum group grouping: grouping the virtual customers according to temperature compatibility; S4: Neural combination path planning: planning an initial path for each group by using a pre-trained neural combination optimization model; S5: Complete path merging: splicing the initial paths of each group to form a complete initial scheme; S6: Large neighborhood search optimization: optimizing the complete initial scheme by using an improved operator to output an optimal distribution path.

[0044] The perishable food distribution method based on the neural combination optimization model of the present embodiment has the same advantages as the perishable food distribution system based on the neural combination optimization model described above, and will not be described here.

[0045] In a specific example, the temperature compatibility graph construction method of the maximum group grouping in step S3 is: initializing an undirected graph G = (V, E), V is a virtual customer set, and if the temperature range intersection of virtual customers i and j is not empty, then an edge (i, j) is added between i and j to the edge set E.

[0046] In a specific example, the improved operator optimization in step S6 includes calculating a benchmark cost, and the benchmark cost calculation formula is Cost = w1 x total cost + w2 x total mileage + w3 x number of refrigerated vehicles + w4 x mass loss during transportation, where w1, w2, w3 and w4 are weight coefficients. , and , T is a temperature parameter of the current iteration number (T = T0 x (1 - l)k), and T0 is the initial temperature parameter. ​​​​​, where k is the current iteration round, k∈[1,200].

[0047] In the specific implementation: Data simulation preprocessing: For a fresh food delivery area in a certain city, the locations of 50 customers (x) were collected. i ,y i Delivery time window (e.g., 7:00-12:00), required items (lemons, frozen fruit, etc.), item temperature requirements (lemons 10-15℃, frozen fruit -30 to -15℃), and quality loss parameters (α=0.02 / h). =0.1 / h), after cleaning the abnormal data, a structured dataset is formed.

[0048] Virtual customer segmentation: Customer A (location (30,40), time window 7:00-12:00) who needs to buy lemons and frozen fruit at the same time is split into virtual customers A1 (demand for lemons 20kg, 10-15℃) and A2 (demand for frozen fruit 42kg, -30 to -15℃).

[0049] Maximum Clique Grouping: Construct a temperature compatibility graph containing 80 virtual customers and divide them into 4 groups using the maximum clique algorithm. A1 is grouped with other virtual customers with a demand of 10-15℃, and A1 is grouped with other virtual customers with a demand of -30 to -15℃.

[0050] Neural synthesis path planning: Call the pre-trained model (number of attention heads = 4, hidden layer dimension = 128) to generate paths for 4 groups respectively, such as the first group path [0,2,7,3,0,5,6,9,0].

[0051] Complete path merging: merge the 4 groups of paths into the initial complete path [0,2,7,3,0,5,6,9,0,11,1,4,0,2,8,10,0,15,13,0,12,16,0].

[0052] Large neighborhood search optimization: Let =0.4, =0.3, =0.3, =100, λ=0.95, after 200 iterations, the total path mileage is reduced by 12%, the number of refrigerated trucks used is reduced by 1, and the total mass loss is reduced by 8%, thus obtaining the optimal solution.

[0053] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A perishable food delivery system incorporating a neural combinatorial optimization model, characterized in that, It includes a data simulation preprocessing unit, a virtual customer segmentation unit, a maximum group grouping processing unit, a neural combination path planning unit, a complete path merging unit, and a large neighborhood search optimization unit, with each unit connected sequentially. The data simulation preprocessing unit is used to collect and process customer information, perishable food demand information, and food physical property parameters to form structured basic data. The virtual customer segmentation unit is used to split actual customers with multiple item needs into virtual customers with single item needs; The maximum grouping processing unit is used to group virtual customers according to temperature compatibility. The neural combinatorial path planning unit is used to plan an initial path for each group of virtual customers using a pre-trained neural combinatorial optimization model. The complete path merging unit is used to combine the initial paths of each group to form a complete initial delivery plan; The large neighborhood search optimization unit is used to optimize the complete initial delivery plan by improving the operators.

2. The perishable food delivery system based on the fusion neural combinatorial optimization model according to claim 1, characterized in that, The data simulation preprocessing unit processes data including customer geographic location coordinates ( , ), delivery time range ( , ), types and quantities of goods, minimum tolerable temperature for food transportation Maximum tolerable temperature Quality loss rate during closed-door transport (a) and quality loss rate during open-door handling (b) The processing includes cleaning up abnormal data and interpolating and supplementing missing data.

3. The perishable food delivery system based on the fusion neural combinatorial optimization model according to claim 1, characterized in that, Each virtual customer segmented by the virtual customer segmentation unit retains the original actual customer's geographical location coordinates, delivery time range, and corresponding item temperature requirement range. , ).

4. The perishable food delivery system based on the fusion neural combinatorial optimization model according to claim 1, characterized in that, The grouping constraint of the maximum cluster grouping processing unit is "the intersection of the item temperature requirement ranges of all virtual customers in the same group is not empty". The grouping process includes constructing a temperature compatibility graph, initializing the ungrouped set and the grouped set, iteratively extracting the maximum cluster and updating the set.

5. The perishable food delivery system based on the fusion neural combinatorial optimization model according to claim 1, characterized in that, The neural combinatorial path planning unit's neural combinatorial optimization model adopts an encoder-decoder structure, with the encoder module being a graph neural network (GNN) and the decoder module being a Transformer structure with an attention mechanism. The training objectives are to minimize the total path mileage, the number of refrigerated trucks used, and the food quality loss. Furthermore, the hyperparameters can be automatically adjusted according to the group size.

6. The perishable food delivery system based on the fusion neural combinatorial optimization model according to claim 1, characterized in that, The method for splicing the complete path merging unit is as follows: the first group of paths is used as the base segment, and the starting point of each subsequent group of paths is connected to the ending point of the previous group of paths, without the need for conflict coordination.

7. The perishable food delivery system based on the fusion neural combinatorial optimization model according to claim 1, characterized in that, The large neighborhood search optimization unit takes the complete initial delivery plan as the optimization object, and achieves path depth optimization by improving the collaborative operation of the destruction operator and the repair operator, combined with the simulated annealing acceptance criterion.

8. A method for distributing perishable food using a neural combinatorial optimization model, characterized in that, Includes the following steps: S1: Data simulation preprocessing: Collect and process customer information, perishable food demand information and food physical property parameters to form structured data; S2: Virtual Customer Segmentation: This segmentation breaks down real customers with multiple item needs into virtual customers with single item needs. S3: Maximum Grouping: Group virtual customers by temperature compatibility; S4: Neural Combination Path Planning: Initial paths are planned for each group using a pre-trained neural combination optimization model; S5: Complete Path Merging: Combine the initial paths of each group to form a complete initial solution; S6: Large Neighborhood Search Optimization: Improve the complete initial solution by refining the operators and output the optimal delivery route.

9. The perishable food delivery system based on the fusion neural combinatorial optimization model according to claim 1, characterized in that, In step S3, the temperature compatibility graph of the largest clique group is constructed as follows: initialize an undirected graph G=(V,E), where V is a set of virtual customers. If the intersection of the temperature ranges of virtual customers i and j is not empty, then add an edge (i,j) between i and j to the edge set E.

10. The perishable food delivery system based on the fusion neural combinatorial optimization model according to claim 1, characterized in that, Step S6, improving the operator optimization, includes calculating the baseline cost, where the baseline cost is calculated using the formula Cost = ×Total route mileage+ ×Number of refrigerated vehicles used+ × Quality loss during transportation, including , and Weighting coefficients; Overall cost: The formula for the acceptance probability of simulated annealing is P = T is the temperature parameter for the current iteration number (T= × , where k is the current iteration round, k∈[1,200].

Images