An order prediction-based intelligent logistics scheduling method for plasticizing industry

Abstract

The application discloses a plasticization industry intelligent logistics scheduling method based on order prediction, relates to the plasticization product logistics field, predicts order demands of a plurality of customers for plasticization products within a preset time window based on historical data, and constructs a knowledge graph based on customer nodes, order nodes and model nodes corresponding to the orders; in the knowledge graph, attribute parameters are added to each model node, the attribute parameters including a melt flow rate, a color grade and an additive grade; a switching score function is constructed based on the attribute parameters; under the premise that all orders are completed within the preset time window, a total target function is constructed, the total target function comprehensively considering a total driving distance, a vehicle cleaning cost score and a customer cleaning cost score; and a scheduling scheme is generated by using an optimization algorithm with the minimum total target function as a target. The application effectively solves the cross-contamination problem caused by the traditional scheduling method neglecting model replacement.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of plastic product logistics, more particularly, to a plastic industry intelligent logistics scheduling method based on order prediction. BACKGROUND

[0002] The logistics transportation of the plastic industry mainly adopts vehicle distribution. After the plastic products are delivered to the customers, the customers usually need to store or process the products through devices such as storage tanks, pipelines, hoses and hoppers. These devices are often repeatedly used between multiple orders of the same customer, and therefore, when different plastic products are sequentially unloaded into the same device, residues are easily generated. If the residues are not completely removed, they will enter the storage or processing system of the customer along with the subsequent materials, resulting in cross-contamination.

[0003] The cost of cleaning is closely related to the characteristics of the grades. Color grade is one of the sources of contamination. For example, when the previously unloaded grade is black and the subsequently unloaded grade is transparent or light-colored, even if there is only a small amount of black residue, it will form obvious color mixing in the new product, affecting the quality and even causing the entire batch to be scrapped. Conversely, when switching from transparent to black, the transparent residue has little effect on the final color, and the cleaning requirement and cost are relatively low.

[0004] The difference in additive grade is also important. The larger the additive grade value, the purer it is, and the closer it is to food grade / medical grade. High-grade products have very high requirements for purity, and if there is low-grade residue in the customer's equipment, even in a small proportion, it may cause the newly unloaded high-grade grade to lose its eligibility, resulting in serious losses.

[0005] The difference in melt flow rate not only affects the amount of residue, but also determines the carrying property between different grades. When the previously unloaded grade has a low melt flow rate, the molecular chain of the material is long and is easy to adhere and form residues. However, the subsequently unloaded grade with a high melt flow rate is more likely to carry these residues into the new product, causing significant contamination and requiring additional cleaning. Conversely, when the previously unloaded grade has a high melt flow rate, the amount of residue is small, and when the subsequently unloaded grade has a low melt flow rate, it is not easy to carry the residues of the high melt flow rate, so the degree of contamination is relatively low and the cost of cleaning is relatively low.

[0006] Similarly, when the vehicle transports different grades of products in sequence, it also requires cleaning of the vehicle or unloading auxiliary equipment on the vehicle. Therefore, from the perspective of the vehicle or the unloading of the customer, the sequence of different grades will result in different cleaning costs. If these factors are not fully considered in the scheduling process, it will often lead to decreased transportation efficiency, frequent cleaning and increased costs. However, this problem is ignored in the prior art. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a plasticization industry intelligent logistics scheduling method based on order prediction to solve the problems mentioned in the background art.

[0008] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0009] Based on historical data, the order demand of a plurality of customers for plasticization products within a preset time window is predicted, and a knowledge graph is constructed based on customer nodes and order nodes and order corresponding brand nodes; in the knowledge graph, the customer node points to the order node, indicating that each customer corresponds to an order, and the order node points to the brand node, indicating that each order corresponds to the brand of the plasticization product; the knowledge graph can better and more visually organize the order relationship;

[0010] In the knowledge graph, an attribute parameter is added to each brand node, including melt flow rate, color grade and additive grade;

[0011] A switching score function is constructed based on the attribute parameters, which is used to quantify the vehicle cleaning cost score and customer cleaning cost score when switching from one brand of plasticization product to another brand of plasticization product;

[0012] Under the premise that all orders are completed within the preset time window, a total objective function is constructed, which comprehensively considers the total mileage, vehicle cleaning cost score and customer cleaning cost score;

[0013] An optimization algorithm is used to generate a scheduling scheme with the goal of minimizing the total objective function.

[0014] In some embodiments, the formula of the switching score function is:

[0015]

[0016] Where S(g→h) represents the switching score value from brand g to brand h, and represents the cleaning cost score;

[0017] M(g) and M(h) are the melt flow rate values of brand g and brand h, respectively;

[0018] M max and M min are the maximum and minimum values of the melt flow rate among all brands, respectively;

[0019] L(g) and L(h) are the color grade values of brand g and brand h, respectively, and the larger the value, the lighter the color;

[0020] L max and L min are the maximum and minimum values of the color grade value among all the grades, respectively;

[0021] A(g) and A(h) are the additive grade values of grade g and grade h, respectively, wherein the larger the grade value, the purer the grade;

[0022] A max and A min are the maximum and minimum values of the additive grade value among all the grades, respectively;

[0023] a, b, c are weight coefficients of each item, all in the interval of 0-1, and a+b+c=1;

[0024] max(,) represents the larger value in the parentheses.

[0025] In a simplified embodiment, the melt flow rate value can not be considered, and only the color grade and additive grade are collected for the attribute parameters of the grades, so the formula of the simplified switching score function is:

[0026]

[0027] Further, the formula of the total objective function is:

[0028]

[0029] wherein J is the total objective function;

[0030] D is the driving distance, and λ is the unit mileage cost;

[0031] κ h is the unit score cleaning cost caused by the customer switching the grade of the unloaded material;

[0032] κ v is the unit score cleaning cost caused by the vehicle switching the grade of the transported material;

[0033] is a customer set; is a vehicle set;

[0034] U c is a set of adjacent unloading point grade pairs of customer c;

[0035] R v is a set of adjacent task grade pairs of vehicle v;

[0036] S(g→h) represents the switching score value from grade g to grade h.

[0037] In some embodiments, the process of generating a scheduling scheme based on the total objective function comprises:

[0038] generating an initial feasible solution satisfying basic constraint conditions by using a heuristic construction algorithm, the basic constraint conditions including ensuring that all orders are completed within the preset time window;

[0039] applying a local search operator to iteratively optimize on the basis of the initial feasible solution;

[0040] In the iterative optimization process, the total objective function value is taken as the evaluation index, and when the iteration number reaches a preset threshold or the total objective function converges, the corresponding scheduling scheme is output as the final scheme.

[0041] In some embodiments, the heuristic construction algorithm includes any one of a greedy algorithm, a nearest neighbor algorithm, or a rule-based sequential construction method, and the local search operator includes any one of an exchange operator, an insertion operator, or a two-point reversal operator.

[0042] In some embodiments, the step of predicting the order demand of the plurality of customers for the plasticized product within the preset time window based on historical data includes:

[0043] converting historical order data of each customer into a training sample, the training sample including time features, order quantity features, and grade attribute features;

[0044] training the training sample using a machine learning model;

[0045] inputting customer feature data of a current period based on the trained model, and outputting order demand prediction results of each customer within the preset time window.

[0046] In some embodiments, the machine learning model is any one of a random forest, a gradient boosting tree, or a long short-term memory network.

[0047] In some embodiments, the method further includes, before scheduling, further considering the last transported grade information g' of the vehicle, calculating a switching score value based on the last transported grade g' of the vehicle and the first allocated order grade h' by switching a score function, and further calculating an initial vehicle cleaning cost:

[0048] κ h S(g'→h');

[0049] and adding it to the total objective function.

[0050] In some embodiments, if the vehicle has been cleaned before scheduling, the initial vehicle cleaning cost is set to zero.

[0051] In some embodiments, the method further includes:

[0052] For any vehicle last transported plate number g', combined with the target plate number h in all orders, use the formula ∑ h S(g'→h) to calculate the cleaning score of the vehicle and sort it, and select the top N vehicles for the initial cleaning operation.

[0053] The advantages of the present application over the prior art are that the intelligent logistics scheduling method proposed by the present application quantitatively models the cleaning cost generated by switching between different plate numbers in the plastic industry and unifies it into scheduling optimization, thereby effectively solving the problem of ignoring device residues and cross-contamination in traditional scheduling methods. By establishing attribute parameters such as melt flow rate, color grade, and additive grade, and constructing a switching score function, the present application can truly reflect the cleaning cost required by the vehicle during switching between different plate numbers. For example, for the case of switching from a black plate number to a transparent plate number, the model will give a higher cleaning score, thereby tending to avoid this high-cost switching in optimization. This mechanism can reduce product scrap due to residues and reduce customer complaints and default risk due to contamination. Among them, the present application also organizes orders in the form of a knowledge graph, making the orders more systematic and visual. The present application considers driving mileage, vehicle cleaning cost, and customer cleaning cost in the comprehensive objective function, not only optimizing overall transportation efficiency, but also taking into account the actual problems in the unloading link. At the customer level, since unloading equipment such as storage tanks, hoses, and pipelines are usually shared, the present application also includes the residual risk that may be caused by different unloading sequences in the cost calculation. If the sequence arrangement is reasonable, for example, unloading light-colored or transparent products first and then unloading dark-colored products, the system will automatically identify and reduce the cleaning requirement from the total objective function value; if the sequence arrangement is not proper, the corresponding cleaning score will be increased, forcing the optimization algorithm to adjust the route again, thereby avoiding the loss caused by cross-contamination.

[0054] Further, the present application uses a heuristic construction algorithm to generate an initial scheme, and combines a local search operator for iterative optimization, so that the scheduling result balances between global efficiency and local cleaning cost. By training and predicting historical order data through a machine learning model, customer demand can be more accurately estimated, and temporary cleaning and repeated scheduling caused by unexpected orders can be avoided. In addition, the present application method also introduces vehicle historical transportation information, establishes an initial cleaning cost and vehicle sorting cleaning mechanism, so that when the plate number of the vehicle last transported is significantly different from the new assigned order, cleaning can be arranged at the beginning of scheduling, thereby fundamentally reducing the risk of mid-operation. This consideration makes the scheme more in line with the actual transportation scenario, taking into account both economy and safety. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a schematic diagram of the overall structure of the present application;

[0056] Figure 2is a schematic diagram of constructing a switching score function of the present application.

[0057] Figure 3 is a schematic diagram of constructing a total target function of the present application. DETAILED DESCRIPTION

[0058] The specific embodiments of the present application are described below in conjunction with the accompanying drawings.

[0059] The present application proposes a plasticization industry intelligent logistics scheduling method based on order prediction, which is used for optimizing the distribution process of plasticization products, and reducing the overall logistics cost by predicting customer demand and considering the cleaning cost caused by grade switching.

[0060] As Figure 1 is a general schematic diagram of the present application, the method of the present application comprises the following steps:

[0061] Based on historical data, the order demand of a plurality of customers for plasticization products within a preset time window is predicted, and a knowledge graph is constructed based on customer nodes and order nodes and grade nodes corresponding to the orders; in the knowledge graph, the customer nodes point to the order nodes, the order nodes point to the grade nodes, and the grades are the grades of the plasticization products; the order relationship can be better and more visualized by using the knowledge graph;

[0062] In the knowledge graph, an attribute parameter is added to each grade node, and the attribute parameter includes melt flow rate, color grade and additive grade;

[0063] Based on the attribute parameter, a switching score function is constructed, and the switching score function is used to quantify the vehicle cleaning cost score and the customer cleaning cost score corresponding to the switching from the plasticization product of one grade to the plasticization product of another grade;

[0064] Under the premise that all orders are completed within the preset time window, a total target function is constructed, which comprehensively considers the total driving distance, the vehicle cleaning cost score and the customer cleaning cost score;

[0065] An optimization algorithm is used to generate a scheduling scheme with the goal of minimizing the total target function.

[0066] In more specific embodiments, the method of the present application first predicts the order demand of a plurality of customers for plasticized products within a preset time window based on historical data. The preset time window can be one week or one month in the future, which is flexibly set according to industry demand. To achieve the prediction, the historical order data of each customer is converted into training samples, which include time features such as order date, day of the week, seasonal factors, order quantity features such as historical average quantity, peak quantity, and grade, which can include attribute features such as melt flow rate, color grade, and additive grade. A machine learning model is used to train these training samples.

[0067] The machine learning model can use any of random forest, gradient boosting tree, or long short-term memory network.

[0068] Taking the random forest model as an example, its architecture is composed of multiple decision trees, each tree is constructed by bootstrapping random sampling from the training samples, and a subset of features is randomly selected at each node for splitting to reduce overfitting. The training process involves constructing hundreds of trees, each of which independently learns sample patterns, and finally predicts the order demand through majority voting or average.

[0069] The gradient boosting tree model uses a sequential approach, where each subsequent tree corrects the residual error of the previous tree. The architecture includes weak learners such as shallow decision trees, which are optimized through gradient descent to minimize loss functions such as mean squared error. During training, trees are added step by step until convergence, improving prediction accuracy.

[0070] The long short-term memory network, as a variant of recurrent neural network, is suitable for processing time series data. Its architecture includes input gates, forget gates, and output gates that control information flow through sigmoid and tanh activation functions, and cell states that carry long-term dependencies. Training uses backpropagation through time algorithm to optimize mean squared error loss. Batch size and learning rate are adjusted according to data size, and usually iterate hundreds of rounds until the validation set loss is stable. Based on the trained model, input the customer feature data of the current period, such as recent order trends and external factors, and output the order demand prediction results of each customer within the preset time window, including the predicted order quantity and specific grade demand.

[0071] Next, attribute parameters are established for the grade of plasticized products corresponding to each predicted order, including melt flow rate, color grade, and additive grade. Melt flow rate is expressed in grams per 10 minutes, reflecting material flowability, with higher values indicating stronger flowability. Color grade values indicate lighter colors, with black being low value 1 and transparent being high value 10. Additive grade values indicate higher purity, closer to food-grade / medical-grade, with industrial-grade set to 1 and medical-grade set to 5. These parameters are extracted from the product database or obtained through experimental measurement to ensure accurate reflection of grade characteristics.

[0072] Further, the present application introduces a knowledge graph. Specifically, based on historical data, the order demand of a plurality of customers for plasticized products within a preset time window is predicted, and a knowledge graph composed of customer nodes, order nodes and brand nodes is constructed on this basis, wherein the customer nodes point to the order nodes, and the order nodes point to the corresponding brand nodes, thereby forming a hierarchical relationship of customer-order-brand. In the knowledge graph, each brand node is attached with an attribute parameter, facilitating the rapid calculation of the switching score function. Through the knowledge graph, the order relationship can be better organized and managed in a structured and visualized form, facilitating the analysis of complex order dependencies and brand distribution. The knowledge graph can be implemented through existing graph databases and knowledge modeling tools, such as Neo4j, or open source software based on Python, such as NetworkX, RDFLib, etc.

[0073] Further, as shown in Figure 2 , a switching score function needs to be constructed to quantify the corresponding vehicle cleaning cost score and customer cleaning cost score when switching from a brand of plasticized product to another brand. The formula of this function is:

[0074]

[0075] wherein S(g→h) represents the switching score value from brand g to brand h, and represents the cleaning cost score;

[0076] M(g) and M(h) are the melt flow rate values of brand g and brand h, respectively;

[0077] M max and M min are the maximum and minimum values of the melt flow rate among all brands, respectively;

[0078] L(g) and L(h) are the color grade values of brand g and brand h, respectively, and the larger the value, the lighter the color;

[0079] L max and L min are the maximum and minimum values of the color grade among all brands, respectively;

[0080] A(g) and A(h) are the additive grade values of brand g and brand h, respectively, wherein the larger the grade value, the closer to the food grade;

[0081] A max and A min are the maximum and minimum values of the additive grade among all brands, respectively;

[0082] a, b, c are weight coefficients of each item, all in the interval of 0-1, and a+b+c=1, set independently according to industrial experience, for example, if color pollution is more sensitive, b is larger;

[0083] max(,) represents the larger value in the parentheses, which ensures that only the positive difference is considered, that is, when switching from low to high, the cost is generated, such as the cleaning cost caused by switching from black to transparent color, and when switching from high to low, it is set to zero, such as not considering the cleaning cost of switching from transparent color to black. Although this does not mean that there is no need to clean from high to low, but setting the case from high to low to zero can greatly reduce the calculation amount, thereby greatly reducing the calculation amount of the optimization algorithm. From the perspective of the optimization algorithm, the focus of the optimization algorithm is to find the best strategy rather than to calculate the actual cleaning cost, so the form of max(,) function is sufficient and very effective.

[0084] Further, as shown in Figure 3 , under the premise of ensuring that all orders are completed within the preset time window, the total objective function is constructed, which comprehensively considers the total driving mileage, vehicle cleaning cost score and customer cleaning cost score. Its formula is:

[0085]

[0086] Wherein, J is the total objective function;

[0087] D is the driving mileage, and λ is the unit mileage cost;

[0088] κ h is the unit score cleaning cost caused by the customer switching the model to unload;

[0089] κ v is the unit score cleaning cost caused by the vehicle switching the model to transport;

[0090] is the customer set; is the vehicle set;

[0091] U c is the adjacent unloading point model pair set of customer c; Its meaning is: when a customer unloads multiple orders, the model of the previous order and the model of the subsequent order form an adjacent unloading point model pair. For example, a customer needs to unload products of model g first and then unload products of model h in the same time window, then g→h belongs to the adjacent unloading point model pair of the customer. By counting the model switching between all adjacent unloading points, the set U c is formed.

[0092] R vR is the set of adjacent task number pairs for vehicle v; its meaning is: in the whole delivery task sequence of the same vehicle, the number of the previous task transported and the number of the next task transported form an adjacent task number pair. For example, after the vehicle completes the transportation of the order of number g, the next task is to transport the order of number h, then g→h belongs to the adjacent task number pair of the vehicle. By counting the number of the number switching in the task sequence of the vehicle, the set R is formed v .

[0093] S(g→h) represents the switching score value from number g to number h.

[0094] In practical applications, κ h and κ v The parameters can be determined by experiments, such as selecting typical number switching scenarios, conducting multiple cleaning tests, and measuring the change in average cleaning cost caused by unit score change.

[0095] Further, an optimization algorithm is used to generate a scheduling scheme with the goal of minimizing the total objective function. More specifically, an initial feasible scheduling scheme needs to be generated according to the order prediction results and customer demand constraints. The construction of this scheme usually uses a heuristic method to quickly generate a solution that satisfies the basic constraint conditions, providing a starting point for subsequent optimization. The basic constraint conditions mainly refer to the fact that all orders must be completed within the preset time window, while ensuring that the vehicle capacity and running path do not exceed the actual transportation limits. In order to achieve this goal, various heuristic construction strategies can be used, such as a construction method based on the greedy idea, which gradually selects the currently optimal delivery task to form an initial solution; a nearest neighbor method can also be used to connect customers with close distances to form a delivery path; an order construction method based on experience rules can also be used to arrange different customer orders in the vehicle task sequence according to priority rules.

[0096] After generating the initial scheme, the present application further introduces a local search operator on this basis to fine-tune the path and task sequence and improve the overall scheduling result. The local search process is an iterative optimization mechanism, which constantly tries to find a better solution within the neighborhood of the current solution, and decides whether to accept the improvement according to the change in the objective function value. Common local search operators include the exchange operator, which optimizes the path by exchanging the service order of two customers; the insertion operator, which reduces the total cost by inserting a customer's task into a suitable position in another path; and the two-point reversal operator, which reverses the access order of a section of the path to shorten the travel distance or reduce the switching cost.

[0097] In the iteration process, the total objective function value is used as an evaluation index to measure the pros and cons of the current scheduling scheme. When the system reaches the preset upper limit of the number of iterations, or the objective function value tends to converge in multiple iterations and there is no significant improvement, the optimization process is terminated. At this time, the current obtained scheduling scheme is regarded as the final scheme and output as the actual execution plan of the vehicle and the customer within the specified time window. Through this process of first generating a feasible solution using a heuristic method and then gradually optimizing in a local range, the present application can obtain a scheduling result that meets the constraint conditions and is close to the optimal result within a reasonable calculation time.

[0098] In a simplified scenario, assume that a vehicle needs to deliver goods for four customers in the same time window, and the colors are C1 white, C2 light gray, C3 dark gray, and C4 black. The initial order given by the heuristic construction based on geographical proximity is: C3 dark gray → C1 white → C2 light gray → C4 black. In this order, there is a deep-to-light switch (dark gray to white) at the beginning, which is relatively high in cleaning cost.

[0099] First, use the exchange operator to swap C3 dark gray and C1 white, and get: C1 white → C3 dark gray → C2 light gray → C4 black. In this way, the first segment becomes light to deep, the risk is reduced, but there is still a deep-to-light (dark gray to light gray) in the middle.

[0100] Then use the two-point reversal operator to reverse the middle segment "C3 dark gray → C2 light gray" as a whole to "C2 light gray → C3 dark gray". The sequence becomes: C1 white → C2 light gray → C3 dark gray → C4 black. At this time, the path has realized a monotonic transition from lightest to darkest, the cleaning cost of color switching is minimized, and the original high-risk switch is eliminated.

[0101] In order to verify the local optimum, try the insertion operator again, for example, insert C4 black into the second position to form: C1 white → C4 black → C2 light gray → C3 dark gray. This will introduce an extreme depth change from white to black, and the cleaning cost will obviously increase, so the algorithm will be rolled back and not adopt the modification.

[0102] After the above steps, the algorithm converges to the order "white → light gray → dark gray → black" under the premise of acceptable driving distance. This process logically embodies the functions of the three commonly used operators: exchange is used to quickly eliminate the most obvious high-risk position, two-point reversal is used to adjust a segment of order as a whole to a smoother light-to-dark, and insertion is used to explore a possibly shorter path but will be screened out by the cleaning cost constraint. The final scheme takes into account the feasibility and lower cleaning cost.

[0103] In further embodiments, before scheduling, the present application also considers the last transported license plate information g' of the vehicle, calculates the switching score value based on the switching score function of the last transported license plate g' of the vehicle and the license plate h' of the initial assigned order, and further calculates the initial vehicle cleaning cost κh S(g'→h') and incorporated into the total objective function. This fills the logical gap of the dispatch starting state, ensuring that historical residue impact is included in the optimization. If the vehicle has already experienced a wash before dispatch, the initial vehicle wash cost is set to zero, as there is no residue pollution after washing.

[0104] Further, the method of the present application also includes, for any vehicle last transported with a brand g', combining all target brands h in the orders, using the formula h S(g'→h) to calculate the wash score of the vehicle and sort it, selecting the top N vehicles to perform the initial wash operation, for example, N is one-third of the total number of vehicles. This prioritizes the washing of vehicles with high potential pollution, reducing the overall initial cost.

[0105] Through the above steps, the method of the present application realizes intelligent scheduling of plastic industry logistics. Not only does it predict demand to plan ahead, but it also quantifies the washing motivation through detailed brand attributes and switching functions, ensuring that the scheduling scheme balances between mileage and washing cost. The training of the prediction model expands the use of historical data, and the iteration of the optimization algorithm provides a practical implementation scheme, making the entire process more efficient and reliable. In practical applications, weights and thresholds can be adjusted according to specific industry data to further improve adaptability.

[0106] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art, according to the technical solution and inventive concept of the present application, can make equivalent replacements or changes within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for order prediction-based intelligent logistics scheduling in the plastic industry, characterized in that, The method comprises the following steps: predicting order demands of a plurality of customers for plasticized products within a preset time window based on historical data, and constructing a knowledge graph based on customer nodes and order nodes and model nodes corresponding to the orders; adding attribute parameters to each model node in the knowledge graph, the attribute parameters including melt flow rate, color grade and additive grade; constructing a switching score function based on the attribute parameters, the switching score function being used to quantify vehicle cleaning cost scores and customer cleaning cost scores corresponding to switching from plasticized products of one model to plasticized products of another model; constructing a total target function under the premise that all orders are completed within the preset time window, the total target function taking into account total mileage, vehicle cleaning cost scores and customer cleaning cost scores; generating a scheduling scheme by using an optimization algorithm with the aim of minimizing the total target function; a formula of the switching score function is: ； wherein, represents a switching score value from the grade to the grade of the cleaning cost score; and are melt flow rate values of grades , grade , respectively; and MFRmax and MFRmin are the maximum and minimum values of the melt flow rate among all the grades, respectively. and Color Grade values for grades , and , respectively, with higher values indicating lighter color; and respectively, the maximum and minimum values of the color grade values in all the brands; and are additive grade values for grades , grade , respectively, where the greater the grade value, the purer the grade; and respectively the maximum and minimum values of the additive rating in all the grades. , , are weight coefficients of each item, all in the interval of 0-1, and + + =1. denotes the larger of the values in the brackets.

2. The method of claim 1, wherein, a formula of the total target function is: ； wherein, is the total objective function; for the driving range, for the unit range cost; Unit score cleaning cost due to customer switching brands; Unit score washing fees resulting from switching license plates for vehicles; for a customer set; for a vehicle set; To the customer a set of adjacent drop point grades; for a vehicle adjacent task plate number pair sets; denotes a switch to the brand Switch to brand Switch score value. 3.The method of claim 1, wherein, the process of generating the scheduling scheme based on the total target function comprises: generating an initial feasible scheme that meets basic constraint conditions by using a heuristic construction algorithm, the basic constraint conditions including ensuring that all orders are completed within the preset time window; iteratively optimizing the initial feasible scheme by using a local search operator; in the iterative optimization process, taking the total target function value as an evaluation index, outputting a corresponding scheduling scheme as a final scheme when the number of iterations reaches a preset threshold or the total target function converges.

4. The method of claim 3, wherein the method further comprises: The heuristic construction algorithm includes any one of a greedy algorithm, a nearest neighbor algorithm or a rule-based sequential construction method, and the local search operator includes any one of an exchange operator, an insertion operator or a two-point reversal operator. 5.The plastic industry smart logistics scheduling method based on order prediction of claim 1, wherein, The step of predicting order demands of a plurality of customers for plasticized products within a preset time window based on historical data comprises: transforming historical order data of each customer into training samples, the training samples including time features, order quantity features and model attribute features; training the training samples by using a machine learning model; inputting customer feature data of a current period into the trained model to output order demand prediction results of each customer within the preset time window.

6. The method of claim 5, wherein the method further comprises: The machine learning model is any one of a random forest, a gradient boosting tree or a long short-term memory network.

7. The method of claim 2, wherein the method further comprises: The method further comprises: further considering the last transported plate number information of the vehicle before scheduling , based on the last transported plate number of the vehicle and the plate number of the initial allocation order Calculate the switching score value by switching the scoring function, and further calculate the initial vehicle cleaning cost: ； The total target function is also taken into account. 8.The method of claim 7, wherein, If the vehicle has been cleaned before scheduling, the initial vehicle cleaning cost is set to zero. 9.The method of claim 7, wherein, The method further comprises: For any vehicle last transported plate number , combining target plate numbers in all orders , using formula to calculate the cleaning score of the vehicle and rank, select the vehicle with the highest ranking to perform the initial cleaning operation. N ​

Images