Method, system and device for optimizing a freight network for combined road-rail transport with external collection and internal distribution

Abstract

The application provides an optimization method, system and equipment for a public and railway combined transport delivery network under external set internal matching, and belongs to the technical field of logistics distribution network optimization. The method is aimed at the problem that the traditional optimization algorithm in the prior art is prone to local optimization and difficult to obtain a globally satisfactory solution, and uses a deep learning assisted genetic algorithm to solve a pre-constructed public and railway combined transport network model. The algorithm performs real-time sensing on the population evolution state through a deep Q network, dynamically and intelligently selects a genetic operation strategy, and combines an experience playback mechanism to perform self-training and parameter updating, and finally outputs an optimal two-stage public and railway combined transport hub site selection and network freight flow distribution scheme. The application uses the above optimization method, system and equipment for a public and railway combined transport delivery network under external set internal matching, effectively improves the global search capability and adaptive level of the algorithm, and can obtain a public and railway combined transport network optimization scheme which is lower in cost, better in layout, and stable and reliable.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of logistics distribution network optimization, and particularly relates to a method, system and device for optimizing a public-transport-railway combined delivery network under an external collection and internal distribution mode. BACKGROUND

[0002] As the main mode of urban freight transportation, highways have the advantages of wide coverage, high flexibility and high safety, and play an irreplaceable role in the entire distribution system. However, single highway transportation also has problems such as high carbon emissions, small transport capacity and high cost, which directly affect the urban environment system, freight transportation efficiency and the price of living materials. By combining railway and highway transportation, the advantages of large railway transport capacity, small carbon emissions and high flexibility of highway can effectively solve the above problems. By selecting a public-transport-railway combined hub and optimizing the delivery organization process, the comprehensive service level of the public-transport-railway combined transportation can be improved, thereby updating the urban freight transportation system and realizing external collection and internal distribution.

[0003] In the research on external collection and internal distribution of public-transport-railway combined transportation, the traditional method usually adopts the research idea of traditional urban logistics, and uses a traditional heuristic algorithm to build a delivery system among three-level stations. However, the traditional heuristic algorithm is prone to local optimization and difficult to achieve global optimization. Specifically, in the external collection and internal distribution problem, the traditional genetic algorithm has a single strategy in the selection, crossover and mutation processes. The selection operator usually only uses a single method of roulette or tournament, and cannot dynamically adjust according to the population evolution stage. The crossover operator is fixed on a single-point or two-point crossover strategy, and has limited adaptability to different coding structures. The mutation probability and site selection method in the mutation operator are not adaptive, which can easily lead to loss of population diversity or excessive mutation. Further, the algorithm converges to a certain node selection scheme in the early stage, which may be a local optimum rather than a global optimum. This selection may lead to a long overall delivery path, increasing the transportation time and cost. Further, it exacerbates the overall inefficiency of the system. This limitation makes the algorithm unable to meet the requirements of efficient, flexible and low-cost logistics when applied to the research on the end delivery system of external collection and internal distribution of public-transport-railway combined transportation, and may result in a situation where the model does not match the actual situation in practical applications.

[0004] Therefore, there is an urgent need for a new solution to solve the above problems. SUMMARY

[0005] The purpose of the present application is to provide an optimization method, system and device for a public and railway combined transport delivery network under external collection and internal distribution, which solves a pre-constructed public and railway combined transport network model by introducing a deep learning assisted genetic algorithm; the algorithm uses a deep Q network to dynamically select a genetic operation strategy according to the real-time evolution state of the population, and combines an experience replay mechanism for self-training, thereby effectively overcoming the problems of traditional optimization methods, such as being prone to local optimization and strategy solidification, and finally obtaining a public and railway combined transport hub site selection and freight flow distribution scheme with lower cost, better layout and stable reliability.

[0006] To achieve the above-mentioned purpose, the present application provides an optimization method for a public and railway combined transport delivery network under external collection and internal distribution, which is based on a pre-constructed service network including a first-level logistics distribution center, a second-level public and railway combined transport hub and a third-level terminal distribution site, and a multi-objective distribution optimization model with the objective of minimizing the total cost, and the optimization steps of the method include: using a deep learning assisted genetic algorithm to solve the multi-objective distribution optimization model to obtain a site selection and network freight flow distribution scheme for the second-level public and railway combined transport hub; wherein the deep learning assisted genetic algorithm performs the following processes:

[0007] Step S1, encoding the solution based on the node and freight flow relationship of the distribution network, generating an initial population, and calculating the fitness of the population individuals;

[0008] Step S2, calculating a multi-dimensional state vector according to the current population to represent the evolution state of the population;

[0009] Step S3, inputting the multi-dimensional state vector into a deep Q network; the deep Q network outputs Q values of a plurality of candidate strategies for at least one of selection, crossover and mutation operations; comparing the Q values with an exploration threshold parameter to dynamically select the evolution strategy to be executed in the current generation;

[0010] Step S4, executing the selected selection, crossover and mutation strategy, updating the population and calculating the reward value; storing the initial state of this iteration, the selected action, the obtained reward and the state after evolution as an experience data into an experience replay buffer;

[0011] Step S5, sampling a batch of experience data from the experience replay buffer, training the deep Q network through forward propagation and back propagation, and updating the network parameters thereof;

[0012] Step S6, repeating steps S2 to S5 until the termination condition is met, and outputting the final optimal solution as the network optimization scheme.

[0013] Preferably, in step S1, binary encoding is used, and the chromosome encoding includes two sections, respectively representing the distribution relationship between the third-level terminal distribution site and the second-level public and railway combined transport hub, and the supply relationship between the second-level public and railway combined transport hub and the first-level logistics distribution center.

[0014] Preferably, in step S2, the multi-dimensional state vector is a 15-dimensional vector, including: population diversity, average fitness, maximum fitness, normalized minimum cost, normalized average cost, normalized historical optimal cost, current generation ratio, stagnation generation ratio, exploration rate, chromosome segment standard deviation minimum value, chromosome segment standard deviation maximum value, chromosome segment standard deviation average value, similarity to the optimal individual, relative cost range, relative size of cost difference.

[0015] Preferably, in step S3, the deep Q network includes an input layer, at least one hidden layer, and an output layer.

[0016] Preferably, in step S3, the exploration threshold parameter is epsilon, and a random number is generated when making a decision, and if the random number is less than epsilon, a random action is selected, otherwise the action with the maximum Q value is selected.

[0017] Preferably, in step S3, the candidate strategies provided for the selection operation include: roulette wheel selection, tournament selection, elite preservation combined tournament selection, linear ranking selection, diversity and fitness balanced selection.

[0018] And / or, the candidate strategies provided for the crossover operation include: single-point crossover, two-point crossover, cyclic crossover, partial mapping crossover, sequential crossover.

[0019] And / or, the candidate strategies provided for the mutation operation include: swap mutation, insertion mutation, reversal mutation, perturbation mutation, adaptive swap mutation.

[0020] Preferably, in step S4, the reward value calculation formula is as follows:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] Wherein, represents the reward value, represents the reward brought by the selection operation, represents the reward brought by the crossover operation, represents the reward brought by the mutation operation, represents the average cost of the population before the selection stage, represents the average cost of the population after the selection stage is executed, represents the fixed reward.

[0026] Preferably, in step S6, the termination condition is that the number of evolutionary iterations reaches a preset value.

[0027] The application also provides an optimization system for a public-transport combined delivery network under external collection and internal distribution, comprising:

[0028] A network model construction module is configured to construct a service network comprising a first-level logistics distribution center, a second-level public-transport combined hub and a third-level terminal distribution site, and a multi-objective distribution optimization model with the objective of minimizing total cost;

[0029] An encoding and population initialization module is configured to encode solutions based on the relationship between nodes of the distribution network and freight flows, generate an initial population, and calculate the fitness of population individuals;

[0030] A state feature extraction module is configured to calculate a multi-dimensional state vector according to the current population to represent the evolutionary state of the population;

[0031] A strategy decision module is configured to input the multi-dimensional state vector into a deep Q network, output Q values of a plurality of candidate strategies for at least one of selection, crossover and mutation operations from the deep Q network, and compare the Q values with an exploration threshold parameter to dynamically select an evolutionary strategy to be executed in the current generation;

[0032] A population evolution and experience storage module is configured to execute the selected selection, crossover and mutation strategy, update the population and calculate a reward value, and store the initial state, selected action, obtained reward and evolved state of the current iteration as experience data in an experience replay buffer;

[0033] A network training module is configured to sample experience data from the experience replay buffer, train the deep Q network through forward propagation and back propagation, and update network parameters of the deep Q network;

[0034] An iterative optimization module is configured to repeatedly call the state feature extraction module, the strategy decision module, the population evolution and experience storage module and the network training module until a termination condition is met;

[0035] An output module is configured to output a final second-level public-transport combined hub site selection and network freight flow allocation scheme.

[0036] The application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the optimization method for a public-transport combined delivery network under external collection and internal distribution when executing the computer program.

[0037] Therefore, the optimization method, system and device for a public-transport combined delivery network under external collection and internal distribution have the following beneficial technical effects:

[0038] (1) In view of the problem that the traditional method is easy to fall into local optimum when optimizing the three-level (distribution center-hub-terminal) public and land combined transport network, leading to improper hub location and detouring distribution path, the dynamic strategy selection mechanism of the deep learning assisted genetic algorithm is used in the application, which significantly enhances the global search ability of the algorithm in complex solution space. This makes the optimization result jump out of the suboptimal node allocation scheme and find a globally optimal network layout with shorter transport distance and more efficient transit, so as to realize the systematic reduction of the total cost of transportation, transit and carbon emission under the premise of meeting all supply and demand constraints.

[0039] (2) In view of the realistic planning problem that needs to balance economic cost and environmental benefit (transportation cost, hub operation cost, carbon emission cost) and is subject to many constraints (supply and demand balance, capacity limitation, flow conservation), the traditional algorithm parameters and strategies are fixed and difficult to adapt. The self-perception and self-learning of the population state by the deep Q network enable the algorithm to dynamically adjust the evolution strategy according to the solving process, intelligently balance the optimization process of different targets, and effectively handle the complex conditions of multiple cargo types and multiple capacity constraints. This improves the adaptability and solving efficiency of the algorithm to real complex scenarios, making the optimization scheme more operable.

[0040] (3) The result of the traditional optimization method is greatly affected by the initial solution and random operation, and the difference between multiple solutions may be significant, which is not conducive to stable decision-making. The experience replay and long-term return evaluation mechanism of reinforcement learning are introduced into the evolution process in the application, which reduces the blind randomness of strategy selection through learning from historical evolution experience and guides the algorithm to converge in a more stable direction. Therefore, consistent and high-quality network optimization schemes can be obtained in different solutions, significantly improving the reliability and decision-making reference value of hub location and freight flow distribution planning results. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The genetic algorithm framework based on deep learning;

[0042] Figure 2 The binary encoding process;

[0043] Figure 3 The cost comparison of the two algorithms and the improvement rate of the combined algorithm compared with the traditional algorithm. DETAILED DESCRIPTION

[0044] The technical solutions of the application are further described below through the drawings and examples.

[0045] Unless otherwise defined, the technical terms or scientific terms used in the application shall have the usual meaning understood by those skilled in the art to which the application belongs.

[0046] Example 1

[0047] The embodiment provides an optimization method of a public-transport-railway combined delivery network with external collection and internal distribution. The core of the method is that, for a service network including a first logistics distribution center (city periphery), a second public-transport-railway combined hub (city interior) and a third terminal distribution site (street) which has been constructed in a traditional way (for example, based on expert evaluation, geographic information system analysis or simple cost model) and a multi-objective mathematical model which has been established to minimize total cost (usually including transportation, transfer, fixed facility and environmental cost) as an objective, a deep learning assisted genetic algorithm is used to perform efficient and global optimization solution. The implementation steps of the method are described in detail below with reference to the flowchart shown in Fig. 1. Figure 1

[0048] Step S101, constructing a public-transport-railway combined distribution service network and an optimization model framework.

[0049] Firstly, the three-level structure of the network is defined: the first level node is a logistics distribution center (set ) in the city periphery, responsible for receiving and storing goods from the region; the second level node is a public-transport-railway combined hub (set ) to be selected or preliminarily selected in the city interior, responsible for the core functions of railway and highway transportation mode conversion and goods transfer and distribution; the third level node is the final service object, i.e. a terminal distribution site (set ), such as a distribution center in each street. Secondly, according to actual business rules, resource limitations and optimization objectives (such as economy and environmental protection), a corresponding multi-objective mixed integer programming model is established. The model usually includes continuous decision variables for representing the allocation relationship of goods flow, 0-1 decision variables for representing whether the hub is selected, and 0-1 auxiliary decision variables for representing the distribution relationship. The objective function of the model aims to minimize the total cost, which is represented as:

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] The constraint conditions are as follows:

[0055] Constraint condition 1: ;

[0056] Constraint condition 2: ;

[0057] Constraint condition 3: ;

[0058] Constraint condition 4:​ ;

[0059] Constraint 5: ;

[0060] Constraint 6: ;

[0061] Constraint 7: ;

[0062] Constraint 8: ;

[0063] Constraint 9: ;

[0064] Constraint 10: ;

[0065] Constraint 11: ;

[0066] in, This represents the transportation costs of railways and highways. This represents the hub operation and transfer costs, including fixed costs (daily operation and maintenance expenses of hub infrastructure) and variable costs (operational costs such as loading, unloading and handling of goods within the hub). It represents the cost of carbon emissions, reflecting the adverse environmental impact of carbon emissions generated by rail transport, road transport, and hub transfer operations; It represents the collection of peripheral logistics distribution centers. Represents a collection of terminal delivery stations. Indicates an index; This represents the set of intermodal transport hubs within a city. Indicates an index; It represents a collection of different kinds of goods. Indicates an index; They represent logistics distribution centers. To alternative rail-road intermodal transport hubs The railway transport distance, the road transport distance from the alternative rail-road intermodal transport transfer node v to the terminal distribution station q; , These respectively represent the number of times the railway commuter train transports the first... Unit cost of goods, road freight truck transportation The unit cost of this type of goods; , These respectively represent from the logistics distribution center Transported to alternative nodes of the rail-road intermodal transport hub The The transport volume of this type of cargo, and alternative nodes for multimodal transport hubs. Transportation to terminal distribution sites The first class of goods transported by the transportation; , respectively represent the fixed operating cost of the first alternative rail-road intermodal transfer node to carry out intermodal transport business, and the unit cost of the second alternative rail-road intermodal transfer node for the transfer of goods unloading; represents a 0-1 decision variable, if the rail-road intermodal hub alternative node v is selected, ; otherwise, ; represents the carbon emission tax rate; represents the carbon emission intensity per unit of goods transfer; , respectively represent the carbon emission intensity of the railway shuttle train transportation and the carbon emission intensity of the highway truck transportation; , respectively represent the total capacity of the first rail-road intermodal transfer node for the transfer of various types of goods, and the total capacity of the second rail-road intermodal node for the transfer of the first class of goods; represents the total supply of the first logistics distribution center; represents the demand of the first class of goods of the qth terminal distribution site; represents a 0-1 decision variable, if the qth terminal distribution site is distributed by the vth rail-road intermodal hub alternative node, ; otherwise, .

[0067] S102, the model is solved by using deep learning assisted genetic algorithm.

[0068] Referring to Figure 1 , the algorithm embeds the decision-making ability of deep reinforcement learning (DRL) into the evolution framework of genetic algorithm (GA), forming an intelligent optimizer that can self-learn and adaptively adjust the evolution strategy. The specific execution process is as follows:

[0069] S1021, encoding and population initialization. Design the chromosome coding scheme according to the decision variable structure of the model in step S101. A typical implementation is to use binary coding (such as Figure 2 ). For example, the chromosome can be divided into two main sections: the first section encodes the allocation relationship between the third terminal distribution site and the second rail-road intermodal hub (the length can be designed as: ), and the second section encodes the supply relationship between the second rail-road intermodal hub and the first logistics distribution center (the length can be designed as: The total encoded length of a chromosome is defined as An initial population of chromosomes is randomly generated containing chromosomes of this kind and the fitness of each chromosome (individual) is calculated according to the objective function of the model. represents the th individual in the population, whose fitness is calculated by the function . The objective function value of this individual is and represent the maximum and minimum of the objective function in the current population, respectively. The objective function value is converted to the fitness function value using the following formula:

[0070] ;

[0071] where is a smoothing factor to avoid invalid values and prevent the fitness function denominator from being zero, which leads to invalid function when is equal to . In this embodiment takes 0.01.

[0072] S1022, population state feature extraction. At the beginning of each generation evolution, statistical analysis is performed on the current population to calculate a multi-dimensional state vector to comprehensively characterize the health status and evolution stage of the population. The state vector is the basis for the decision of the deep Q network. A specific and effective implementation is to construct a 15-dimensional state vector, the dimensions of which can include:

[0073] Population diversity :

[0074] ;

[0075] Average fitness :

[0076] ;

[0077] Maximum fitness :

[0078] ;

[0079] Normalized minimum cost :

[0080] ;

[0081] Normalized average cost :

[0082] ;

[0083] Normalized historical best cost :

[0084] ;

[0085] Current algebraic proportion :

[0086] ;

[0087] Stagnation algebraic proportion :

[0088] ;

[0089] Exploration rate :

[0090] ;

[0091] Chromosome segment standard deviation minimum :

[0092] ;

[0093] Chromosome segment standard deviation maximum :

[0094] ;

[0095] Chromosome segment standard deviation average :

[0096] ;

[0097] Similarity to best individual :

[0098] ;

[0099] Relative cost range :

[0100] ;

[0101] Relative size of cost difference :

[0102] ;

[0103] wherein, denotes chromosome length; denotes the th gene of the th chromosome; denotes the average of the gene positions; fitness of the chromosome; current generation number; number of generations without improvement; initial exploration rate; decay rate; average of the gene in the th segment; population size (total number of individuals); fitness value of the th chromosome; global best fitness value among all the iterations; maximum number of iterations of the genetic algorithm; value of the th gene of the best chromosome in history.

[0104] S1023, Evolutionary strategy decision based on deep Q network. Input the state vector into a pre-constructed deep Q network (DQN). The DQN usually contains an input layer (the number of neurons is equal to the dimension of the state vector), at least one hidden layer (such as 32 neurons), and an output layer. The output layer outputs a set of Q values for each of the three key operations of the genetic algorithm (selection, crossover, and mutation), each Q value corresponding to a specific operation strategy (action). For example:

[0105] Selection action set: may include {roulette selection, tournament selection, elite preservation combined with tournament selection, linear ranking selection, diversity and fitness balanced selection}.

[0106] Crossover action set: may include {single-point crossover, two-point crossover, cyclic crossover, partial mapping crossover, sequential crossover}.

[0107] Mutation action set: may include {swap mutation, insertion mutation, reversal mutation, perturbation mutation, adaptive swap mutation}.

[0108] When making a decision, the system generates a random number rand, which is compared with the dynamically decaying exploration rate epsilon. If rand<epsilon, a random action is selected from the selection action set, the crossover action set, and the mutation action set (exploration); otherwise, the action corresponding to the highest Q value is selected.

[0109] S1024, Population evolution and experience storage. Perform the selected action combination in the previous step, and perform the corresponding selection, crossover, and mutation operations on the population to obtain a new generation of population .

[0110] Reward Value The calculation includes two parts: average cost improvement and improvement through finding a new global optimum. The first three formulas are for the average cost improvement part, which involves the selection of ( ),cross( ),Mutations( (Three stages) That is, the average cost of the population before the selection phase. This represents the average population cost after the selection phase is executed.

[0111] An improvement to finding a new global optimum is that if a new global optimum is found, breaking the historical record, a fixed reward will be given. It is evenly distributed among the three stages.

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] Subsequently, the experience tuples of this decision are stored in a fixed-size experience replay buffer D.

[0118] S1025, Deep Q-network training. Periodically replaying data from the experience buffer. A small batch (e.g., batch_size=32) of empirical samples is randomly sampled. These samples are then used to train the DQN. The training process employs a standard deep Q-learning algorithm: the predicted values ​​of the current Q-network are calculated through forward propagation, the target Q-value is calculated through the target network (or a fixed target), the mean squared error (MSE) is used as the loss function, and the parameters of the current Q-network are updated through backpropagation and gradient descent algorithms (e.g., the Adam optimizer). Simultaneously, the exploration rate is decayed according to a predetermined strategy. ,in, This indicates the updated exploration rate.

[0119] S1026, Iterative Optimization and Output. Repeat steps S1022 to S1025 until a preset termination condition is met, such as reaching the maximum number of generations (e.g., 800 generations) or fitness no longer improving over multiple consecutive generations. Finally, output the individual with the highest fitness in each generation of the population. Decoding this individual yields the optimal secondary rail-road intermodal transport hub site selection scheme and a detailed cargo flow allocation scheme. , (value).

[0120] S103, apply the optimization scheme.

[0121] The network optimization scheme output by step S1026 is applied to actual logistics planning and operation to guide the site selection of the combined transport hub, the allocation of transportation resources, and the arrangement of daily distribution paths, so as to realize the comprehensive goals of reducing total cost, improving efficiency, and reducing carbon emissions.

[0122] The embodiment deeply fuses the adaptive decision-making ability of deep reinforcement learning and the group search ability of genetic algorithm, and constructs an intelligent solver that can dynamically adjust the evolution strategy according to the real-time state. It effectively overcomes the defects of single strategy, fixed parameters and easy prematurity of traditional genetic algorithm, and exhibits stronger global search ability, faster convergence speed and more stable solution quality when solving complex combined transport network optimization problems.

[0123] The application will be further described below through specific examples.

[0124] (1) Network and model parameter setting.

[0125] Network node determination:

[0126] Primary logistics distribution center: Select A logistics base and B logistics base in a certain area.

[0127] Secondary combined transport hub: First, 27 railway freight sites with railway freight conditions in the urban area of a certain area are initially selected as candidate nodes. Then, based on 16 evaluation indexes in four categories of basic conditions, operation level, coordination degree and expansion capacity, the weights of each index are determined by using the analytic hierarchy process (AHP), and the TOPSIS comprehensive evaluation method is used to sort the candidate nodes, and finally 18 railway freight sites with the highest comprehensive score are selected as combined transport hubs.

[0128] Third terminal distribution site: All 134 streets under the jurisdiction of the urban area of a certain area are used as the end distribution network.

[0129] Cargo type and demand prediction:

[0130] Cargo type: Two main living materials are considered: agricultural and sideline products (P1) and other living materials (P2).

[0131] Demand prediction: Based on the population base of each street obtained from the seventh national census (2020), combined with the historical population change rate data from 2010 to 2023, the permanent population of each street in 2027 is predicted. According to the per capita demand for living materials predicted by the macro development plan of a certain area, combined with the predicted population of each street, the detailed demand for two types of materials in each street in 2027 is calculated.

[0132] Optimization scenario design: To simulate the application of the "external collection and internal distribution" mode at different development stages, five progressive scenarios are set, in which the proportion of the total transport volume of living materials undertaken by the mode in the urban six districts is 20%, 40%, 60%, 80%, and 100%, respectively. Based on this proportion, the demand of each street is converted as a constraint condition for the input of the model.

[0133] (2) Model solution and result analysis.

[0134] Compared with the traditional genetic algorithm, the deep learning assisted genetic algorithm of the application dynamically selects the most suitable genetic operation strategy (intelligent decision from 5 selection, crossover, and mutation methods) for each generation according to the real-time state of the population (such as diversity, fitness distribution, convergence, etc. 15-dimensional features). At the same time, the maximum long-term reward is set to learn the strategy. The network optimization model under the above five traffic scenarios is solved. The algorithm aims to minimize the total cost, and iteratively optimizes under the premise of meeting all supply and demand, capacity, and flow balance constraints.

[0135] (3) Optimization result example.

[0136] Taking the scenario of "external collection and internal distribution" undertaking 100% of the distribution of living materials as an example, part of the optimal network planning scheme output by the algorithm is shown in Table 1. Table 1 clearly shows which tertiary end stations (streets) each enabled secondary public and truck combined transport hub (a total of 18) is responsible for distributing, and which type of goods is distributed. Table 2 compares the costs of the two algorithms. The deep learning assisted genetic algorithm of the application saves an average of 19.1% in cost compared to the traditional genetic algorithm.

[0137] Table 1 System node when external collection and internal distribution undertake 100% of the distribution of urban living materials

[0138]

[0139] Table 2 Cost comparison of two algorithms

[0140]

[0141] For example, Figure 3As shown, in the five traffic volume scenarios of 20% to 100%, the total cost obtained by the DLR-GA combined algorithm is lower than that of the traditional genetic algorithm, with an average reduction of about 19.1%, and the cost advantage of the DLR-GA combined algorithm is the largest when 80% of the city freight transportation is undertaken by the outer set within the matching. With the increase of traffic volume, the two cost curves rise synchronously, and the DRL-GA still maintains a good optimization advantage in large traffic volume.

[0142] Example Two

[0143] The optimization system of the intermodal freight network under the outer set within the matching includes:

[0144] The network model construction module is configured to construct a service network including a first-level logistics distribution center, a second-level intermodal hub and a third-level terminal distribution site, and a multi-objective distribution optimization model with the objective of minimizing the total cost.

[0145] The encoding and population initialization module is configured to encode solutions based on the relationship between nodes of the distribution network and freight flows, generate an initial population, and calculate the fitness of population individuals.

[0146] The state feature extraction module is configured to calculate a multi-dimensional state vector according to the current population to represent the evolution state of the population.

[0147] The strategy decision module is configured to input the multi-dimensional state vector into the deep Q network, output Q values of a plurality of candidate strategies for at least one of the selection, crossover and mutation operations from the deep Q network, and compare the Q values with an exploration threshold parameter to dynamically select an evolution strategy to be executed in the current generation.

[0148] The population evolution and experience storage module is configured to execute the selected selection, crossover and mutation strategies, update the population and calculate the reward value, and store the initial state of this iteration, the selected action, the obtained reward and the evolved state as experience data in an experience replay buffer.

[0149] The network training module is configured to sample experience data from the experience replay buffer, train the deep Q network through forward propagation and back propagation, and update the network parameters of the deep Q network.

[0150] The iterative optimization module is configured to repeatedly call the state feature extraction module, the strategy decision module, the population evolution and experience storage module, and the network training module until a termination condition is met.

[0151] The output module is configured to output the final second-level intermodal hub location and network freight flow allocation scheme.

[0152] If the above functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts of the prior art that make contributions or parts of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0153] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a list of executable instructions for implementing logic functions, which can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus or device, such as a computer-based system, a system including a processor or other system that can fetch the instructions from the instruction execution system, apparatus or device and execute the instructions, or in conjunction with these instructions. For the purpose of this specification, the "computer-readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by or in connection with an instruction execution system, apparatus or device, or in conjunction with these instructions.

[0154] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection having one or more wires (electrical devices), a portable computer diskette (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CD ROM). In addition, the computer-readable medium can even be paper or other suitable medium on which the program can be printed, as the program can be electronically obtained, for example, by optical scanning of the paper or other medium, followed by editing, interpretation or necessary processing, if necessary, in other suitable ways, and then stored in a computer memory.

[0155] It is worth noting that the contents not elaborated in detail in the present application are all prior art and are well known to those skilled in the art.

[0156] Therefore, the application adopts the above-mentioned optimization method, system and device for the public and private joint transport delivery network, can effectively obtain a globally better public and private joint transport hub site selection and freight flow distribution scheme through an intelligent adaptive optimization mechanism, thereby significantly reducing the comprehensive cost of transportation and operation.

[0157] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for optimizing a freight network of a combined road-rail transport, the method being based on a pre-constructed service network comprising a first-level logistics distribution center, a second-level combined road-rail transport hub, and a third-level terminal distribution site, and a multi-objective distribution optimization model with the objective of minimizing the total cost, characterized in that, The optimization step of the method comprises: solving a multi-objective distribution optimization model by using a deep learning assisted genetic algorithm to obtain a site selection and network freight flow distribution scheme of a two-level public and private joint transport hub; wherein the deep learning assisted genetic algorithm performs the following processes: Step S1, encoding the solution based on the relationship between the nodes of the distribution network and the freight flow, generating an initial population, and calculating the fitness of the population individuals; Step S2, calculating a multi-dimensional state vector according to the current population to represent the evolution state of the population; Step S3, inputting the multi-dimensional state vector into a deep Q network; the deep Q network outputs Q values of a plurality of candidate strategies for at least one of selection, crossover and mutation operations; comparing the Q values with an exploration threshold parameter to dynamically select an evolution strategy to be executed in the current generation; Step S4, executing the selected selection, crossover and mutation strategy, updating the population and calculating the reward value; storing the initial state, the selected action, the obtained reward and the evolved state of this iteration as an experience data into an experience replay buffer; Step S5, sampling a batch of experience data from the experience replay buffer, training the deep Q network through forward propagation and back propagation, and updating the network parameters thereof; Step S6, repeating steps S2 to S5 until a termination condition is met, and outputting the final optimal solution as the network optimization scheme.

2. The optimization method of the outer set and inner set down public transport delivery network according to claim 1, characterized in that, In step S1, binary encoding is used, and the chromosome code contains two sections, which are respectively used to represent the allocation relationship between the three-level terminal distribution station and the two-level public and private joint transport hub, and the supply relationship between the two-level public and private joint transport hub and the one-level logistics distribution center.

3. The optimization method of the outer set and inner set down public transport delivery network according to claim 1, characterized in that, In step S2, the multi-dimensional state vector is a 15-dimensional vector, including: population diversity, average fitness, maximum fitness, normalized minimum cost, normalized average cost, normalized historical optimal cost, current generation ratio, stagnation generation ratio, exploration rate, minimum value of chromosome segment standard deviation, maximum value of chromosome segment standard deviation, average value of chromosome segment standard deviation, similarity to the optimal individual, relative cost range, and relative size of cost difference.

4. The optimization method of the outer set and inner set down public transport delivery network according to claim 1, characterized in that, In step S3, the deep Q network includes an input layer, at least one hidden layer and an output layer.

5. The optimization method of the outer set and inner set down public transport delivery network according to claim 1, characterized in that, In step S3, the exploration threshold parameter is epsilon, and a random number is generated when making a decision. If the random number is less than epsilon, a random action is selected, otherwise the action with the maximum Q value is selected.

6. The optimization method of the outer set and inner set down public transport delivery network according to claim 1, characterized in that, In step S3, the candidate strategies provided for the selection operation include: roulette selection, tournament selection, elite reservation combined tournament selection, linear ranking selection, and diversity and fitness balanced selection; and / or, the candidate strategies provided for the crossover operation include: single-point crossover, two-point crossover, cyclic crossover, partial mapping crossover and sequential crossover; and / or, the candidate strategies provided for the mutation operation include: swap mutation, insertion mutation, reverse mutation, disturbance mutation and adaptive swap mutation.

7. The optimization method of the outer set and inner set down public transport delivery network according to claim 1, characterized in that, In step S4, the reward value is calculated according to the following formula: ； ； ； ； wherein, represents a reward value, represents a reward from a selection operation, represents a reward from a crossover operation, represents a reward from a mutation operation, represents a population average cost before a selection phase, represents a population average cost after a selection phase is performed, represents a fixed reward.

8. The optimization method of the outer set and inner set down public transport delivery network according to claim 1, characterized in that, In step S6, the termination condition is that the number of evolution iterations reaches a preset value.

9. An optimization system for a road-rail intermodal freight network with external collection and internal distribution, characterized in that, including: The network model construction module is configured to construct a service network including a first-level logistics distribution center, a second-level public and private transport hub, and a third-level terminal distribution site, and a multi-objective distribution optimization model with a minimum total cost as an objective; The encoding and population initialization module is configured to encode solutions based on the relationship between nodes of the distribution network and freight flows, generate an initial population, and calculate the fitness of population individuals; The state feature extraction module is configured to calculate a multi-dimensional state vector according to the current population to represent the evolution state of the population; The strategy decision module is configured to input the multi-dimensional state vector into the deep Q network, output Q values of a plurality of candidate strategies for at least one of selection, crossover, and mutation operations from the deep Q network, and compare the Q values with an exploration threshold parameter to dynamically select an evolution strategy to be executed in the current generation; The population evolution and experience storage module is configured to execute the selected selection, crossover, and mutation strategies, update the population, and calculate a reward value, and store the initial state, selected action, obtained reward, and evolved state of the current iteration as experience data in an experience replay buffer; The network training module is configured to sample experience data from the experience replay buffer, train the deep Q network through forward propagation and back propagation, and update network parameters of the deep Q network; The iterative optimization module is configured to repeatedly call the state feature extraction module, the strategy decision module, the population evolution and experience storage module, and the network training module until a termination condition is met; The output module is configured to output a final second-level public and private transport hub site selection and network freight flow allocation scheme.

10. A computer device comprising: A memory and a processor; The memory stores a computer program, and the processor executes the computer program to implement the steps of the optimization method of the public and private transport delivery network under the external collection and internal distribution according to any one of claims 1-8.

Images