Abstract
Recently, fresh agricultural coldchain logistics have been greatly developed with the increasing needs of people’s life. Reducing costs of coldchain distribution has become the main object of loss control in logistics enterprises. The objective of this research is to find a set of optimal routes that minimize the total loss, including fuel cost, refrigeration cost, soft time window penalty cost, and cargo damage cost over transit time. In this paper, the definition and model construction of vehicle routing problem (VRP) with multiobjective minimum lost are introduced first. Then, an ant colony optimization (ACO) algorithm combined with Pareto local search (PLS) is put forward to solve the minimum loss model. In order to avoid the influence of complex road conditions during distribution, the distance matrix and the transit time matrix are both derived from the recommended navigation road based on Emap API. At last, a compare experiment between the traditional method and our proposed method is performed. The results indicate that our method has strong applicability and potential advantages in coldchain logistic and has important practical significance and application value.
1. Introduction
According to relevant media reports, rotting fruit, vegetables, and other foods transported by truck alone are worth about 70 billion yuan each year, causing a huge economic waste. Agricultural products especially for fruit, vegetables, fish, meat, etc. require strictly limited temperature, humidity, and time in the process of transportation and storage. As a branch of the logistics industry, coldchain logistics provides a guarantee for the safe transportation of fresh agricultural products. In the process of transportation from cold storage to customers, a complete coldchain logistics realizes the temperature control of the whole process of refrigerated and frozen food, as well as the closed environment, storage, and transportation during loading and unloading of goods. By selecting the optimal path in the process of coldchain logistics transmission, the circulating rate of fruits and vegetables, meat, and aquatic products is reduced; the waste of resources and the cost of logistics costs are reduced too. The vehicle routing problem (VRP) introduced by Dantzig and Ramser [1] plays a central role in the optimization of distribution networks.
Recently, the loss problem of coldchain distribution has attracted the attention of many scholars and experts. For distribution problem of fresh agricultural products, fuel cost, refrigeration cost, cargo damage cost, and soft time window penalty cost have become the important components of loss control in distribution companies [2]. Li et al. [3] proposed a lowcarbon model for fresh food and a genetic simulated annealing algorithm to solve the model in the coldchain distribution. Yao et al. [4] proposed a minimizing fuel consumption solution to timedependent VRP with time window. Kim et al. [5] proposed a Markov decision process method to solve a dynamic VRP (DVRP) model with nonstationary transit times under actual traffic congestion. Chen et al. [6] developed a hybrid heuristic algorithm including harmony search and neighborhood descent to solve DVRP with time window. Abidi et al. [7] proposed a GA with a simple heuristic to solve a variant of RVPR (Rich VPR) with time widows and dynamically changing orders. Dongdong and Yinzhen [8] proposed a green multitype VRP with time windows to reduce the wastes of fuel consumption and carbon emission and use an improved tabu search algorithm to solve the GMVRPTW. Fan et al. [9] pointed out that in fresh agricultural product coldchain logistics, the total costs are composed of five kinds: fixed, transportation, damage, penalty, and energy consumption. Fang and Ai [10] proposed a mathematical model with soft time window penalty cost, refrigeration cost, cargo damage cost, and a hybrid ant colony algorithm to minimize the total costs. The DVRP mainly deal with the timevariation information of customer demands and road conditions. To share the traffic information and classify traffic conditions, Big Data and classification techniques are used in logistics and transportation. If the drivers receive the information of the traffic congestion or poor weather, they can change their way to reduce the costs and save time. Dimensionality reduction must be performed first in Big Data transmitted and stored process. Thippa et al. [11] proposed a machine learning (ML) algorithm with PCA to reduce the dimension of Big Data when the data sets are high. Gadekallu et al. [12] have studied the hybrid PCAwhale optimization algorithm to extract features and used a deep neural network to classify the diseases of tomato. Guha et al. [13] proposed an ANNbased content classification in combination with grame TFIDF feature descriptor to classify the documents with accurate, sensitive information. The ant colony optimization (ACO) proposed by Colorni et al. [14] has been widely used in solving NPhard vehicle routing problems. Many scholars used ACO algorithms to solve multiobjective combinatorial optimization problems [15–19]. These papers considered the costs of fuel, refrigeration, cargo damage, and delay simultaneously. However, their studies are not comprehensive enough in factors affecting the loss, such as the cost of energy consumption and rotting consumption when the door of compartment is opening.
We construct a total cost model, including soft time window penalty cost, fuel cost, refrigeration cost in transit and during unloading, and cargo damage cost in transit and during unloading. In our model, to take into account the actual traffic condition, the distance matrix and the transit time matrix are both archived from the navigation functions based on EMap API. The traditional heuristic algorithm ACO has some limitations, such as easy stagnation in the initial stage and slow search speed. We perform an ACO algorithm with Pareto local search (PLS) on ants to obtain the uniform Paretooptimal frontier and keep the diversity of Pareto solution set.
The paper is organized as follows. Section 1 introduces the definition and structure of VRP for coldchain logistics. Section 2 constructs a minimum cost model based on actual traffic conditions. Then, an improved ACO algorithm with PLS is presented in Section 3. In Section 4, a compare experiment is performed to indicate our proposed model and the improved method. At last, we make a conclusion of our contributions in Section 5.
2. Problem Description
A coldchain supplier has a warehouse with a certain number of transportation vehicles that delivers a variety of fresh products to a certain number of customers. The capacity of the transportation vehicles is limited, and the vehicles are all the same type with a total capacity at the starting point. The object of vehicle routing problem is to distribute goods to each customer with correct goods, limited time, minimum cost, and so on. VRP with time window means that a vehicle has to visit a customer within a certain time window. So, the service time, the transit time, and the total time must be calculated when dealing with the VRP. The mathematical model of VRPTW in coldchain distribution is defined as follows [20].
There are transportation trucks in this distribution center. The maximum load and the maximum transit time of per vehicle are and T. The th distribution vehicle is responsible for path . is the total customers. The needs and the service time of customer point are and ; means the delivery time from customer points to . is a 01 variable, and means the point in the path . is a 01 variable, and means that the th distribution vehicle travels from to .
3. Model
Minimizing the loss cost is the target of this paper. Therefore, the following paper introduces the structure of the loss cost of cold chain.
3.1. Distance and Transit Time Analysis
In traditional methods, the distance between two points is achieved by Euclidean distance formula and the transit time is achieved by the distance divided by assumed average speed. But, the actual transit time is influenced greatly by road congestion, traffic accidents, traffic control, raining, snowing, etc. Collecting realtime traffic data will result in large amounts of data, and Big Data analysis technology should be considered.
In our literature, the navigation function of Emap is used to avoid the mentioned shortcomings. The distance matrix and the transit time matrix under actual traffic conditions among customer points (including the distribution center) are derived from the Emap public platform [21].
As Figure 1 shows, there are two points with their latitudes and longitudes on the Emap.
As Figure 2 shows, the distance and the transit time between the two points are shown by the navigation function of the Emap platform.
In our study, represents transit time between customer and customer , and is the distance between customer and customer . As abovementioned, and are both achieved from the navigation result based on Emap API.
3.2. Loss Cost Target Construction
3.2.1. Soft Time Window Penalty Cost
Each customer has different numbers of good needs and a soft time window. Penalty cost will be incurred if the vehicle arrives beyond the time window boundary. The service time of each customer point is , is the range of the time window of customer point , is the penalty coefficient when vehicle arrives at customer earlier than , and is the penalty coefficient when vehicle leaves from customer later than . The arrival time of vehicle at customer is . According to the analysis above, the penalty cost of cold chain can be defined as the following:
3.2.2. Fuel Cost
The fuel consumption of the distribution vehicle is inevitable for completing the distribution task. We assume that the fuel cost is proportional to the load of the vehicle pert unit distance. Let be the load of the vehicle and be the load rate of the th vehicle.
Sets and represent no load and full load fuel consumption costs per unit distance of the distribution vehicle. means the fuel consumption cost per unit distance of the th vehicle.
The total cost of fuel consumption can be expressed as
3.2.3. Refrigeration Cost
The temperature of the coldchain logistic must be maintained at a certain low level to keep the freshness of goods. The refrigeration function of vehicles must be operated to achieve the required temperature, which causes the refrigeration cost immediately. (1)Refrigeration cost in transit
The refrigeration cost generated by maintaining the required low temperature per unit time in transit is ; the total cooling cost in transit can be expressed as (2)Refrigeration cost during unloading
Opening the door of the compartment will cause the cool air inside to flow out and the hot air from outside to flow in. To maintain the temperature inside, more energy will be consumed. is set to be the cooling cost per unit time during opening the door, and the total refrigeration cost of unloading is
represents the total cooling cost, which consists of the total cooling cost in transit, and the total cooling cost of unloading can be expressed as
3.2.4. Cargo Damage Cost
The most cargo of the cold chain is fresh goods; with the increase of transit time even in low temperature, the growth of microorganisms will happen. When opening the door of compartment during unloading process, the temperature inside will be unstable which results in the damage of fresh goods more significantly. As mentioned above, is the unit price of goods, and is the load of the th vehicle when it leaves from point . (1)Cost of cargo damage in transit
is given as the loss coefficient of goods per unit weight per unit time in transit; the total cost of cargo damage in transit is as follows: (2)Cost of cargo damage during unloading
During unloading at a customer point, the door of the compartment is open and the cost of cargo damage will increase obviously. Let the cargo damage coefficient of goods per unit weight per unit time when unloading be ; the cost of cargo damage during unloading can be expressed as
The total cost of damage is
In conclusion, the mathematical model of minimum loss target can be expressed as
3.3. Mathematical Model of Minimum Loss Analysis
According to the above analysis, the minimum loss model of the coldchain distribution problem is as follows:
Equation (12) is the objective optimization function which is aimed at finding the minimum total cost during the whole distribution process. Equation (13)–Equation (17) are the constraint conditions, where Equation (13) is a 01 variable; the value 1 means the given customer is serviced by the given vehicle. Equation (14) is a 01 variable too; the value 1 means the given route is travelled by the given vehicle. Equation (15) means each customer point is serviced by one and only one vehicle for distribution. Equation (16) means that the total demand served by the vehicle cannot exceed its maximum load. Equation (17) means that the vehicle can only start or end once.
4. Ant Colony Optimization Algorithm
4.1. Ant Colony Optimization Algorithm
The ant colony optimization (ACO) algorithm with parallelism, positive feedback, and strong robustness is used to solve the NPhard and highly constrained problem. In the 1990s, Italian scholars Dorigo and Maniezzo found that ants will release a certain amount of substance called pheromone in the path when they are searching for food. When they need to choose a path, they prefer to choose the path with high concentration of pheromone. At last, the route with the highest concentration will be selected as an optimist route between their home and the food source.
The ACO algorithm is defined as follows: where is probability of ant transferring from point to point at time ; is the heuristic function, and is the heuristic function important factor; means the pheromone concentration on the path at time , and is the pheromone important factor. represents points that ant is allowed to select in the next step; is the pheromone update over time; is the pheromone factor; and represents the pheromone enhancer.
The rough process of the algorithm is as follows:

4.2. Combining ACO Algorithm with PLS
We applied Pareto local search (PLS) in our paper. After finding a path, PLS is used to further optimize the path. PLS proposed by Paquete et al. is a heuristic algorithm for tackling NPhard multiobjective combinatorial optimization problems in the Pareto sense [22]. Pareto dominance defines a partial order on the set of feasible solutions. The goal when tackling the multiobjective problems in the Pareto sense is to find the set of Paretooptimal solutions. The weak componentwise ordering is used as a mutually nondominated criterion in PLS. Let a solution and its neighborhood . If each in is not dominated by any solution in , is marked visited and is added into . When contains only solutions that have been visited, a Pareto local optimum is achieved [23].
An improved ACO with PLS is proposed in our research to deal with the mode. In our approach, ants are used to optimize solutions and generate new Pareto solutions, and the new pheromone updating strategy is used to control the path selection. The procedure is described as follows:
Step 1. Compute of customers and update the trail level for each .
Step 2. Run PLS, and update the current solution according to calculating results.
Step 3. Compare (Cost, CostBest), and record the costBest and BestSolution.
Step 4. For each move () in BestSolution to update the trail level .
Step 5. Repeat step 1 until the maximum computation time reaches.
The proportion of the pheromone update model of the ACO algorithm with PLS is as follows: where means information entropy value and is transition probability. Equation (19) means that the proportion of the pheromone update will decreases with the increase of the pheromone value. is the positive constant, and its value range is [0,1]. The value 1 means the proportion of the pheromone update is high influenced by the information entropy value, and 0.75 is usually used as the value of . This definition combined a sequence of arithmetic with information entropy to adjust the pheromone adaptively.
5. Experimental Results and Discussion
5.1. Data Set Description
In our experiments, the distribution situation of JingKeLong in Beijing City with one distribution center and 20 customer points is selected to verify our proposed mathematical model and algorithm. The distribution center has 4 vehicles of the same type with an average speed of 60 km/h, a maximum driving time of 1200 minutes, and a maximum load of 2000 kg. Table 1 shows the settings of parameters in our given minimum loss model and ACO+PLS algorithm.
The data of the distribution center and customers are shown in Table 2. The distribution center is defined as 0, and customer points are defined as 1 to 20. Figure 3 shows the latitude and longitude of each point in the system, and Figure 4 shows the actual locations of them in Emap.
Our proposed method needs an actual transit time matrix and distance matrix both derived from Emap API. We supposed that the coordinates of points are stored into vertex [] [2]; the core codes with Emap API to achieve the transit time matrix and distance matrix are described as follows.

Table 3 shows the transit time matrix among customer points and distribution center obtained by Emap API.
Table 4 shows the distance matrix among customer points and distribution center obtained by Emap API.
5.2. Experimental Result and Discussion
In the traditional method, the distance matrix is achieved by the Euclidean distance formula. The Euclidean distance formula of two points on earth is defined as the following: where and are two points on the surface of earth and is the radius of the earth. and represent the longitude and latitude of point .
First, the traditional method ACO with the Euclidean distance matrix is performed to solve the VRP with minimum cost. Figures 5 and 6 are the optimal distribution routes and the trend of cost during iteration processed. The optimal distribution routing scheme, the total transit time, and the total cost based on the traditional method are shown in Table 5.
Second, our proposed method ACO+PLS with the actual transit time matrix and travel distance matrix is performed to solve the VRP with minimum cost. Figures 7 and 8 are the calculated distribution routes on map and the trend of cost during iteration processed. The optimal distribution routing scheme, the total transit time, and the total cost based on the proposed method are shown in Table 6.
From the above compare experimental results, we can see that ACO+PLS with actual traffic conditions can get smaller total cost and save more transit time. Our proposed method can be used to get the optimal solution effectively for VRP of coldchain logistics.
6. Conclusion
This paper studies the model of VPR with minimum cost of coldchain distribution. Minimizing the total cost of distribution can maximize the economic benefit of distribution enterprises. To consider the dynamic change of transit speed under actual traffic conditions, the transit time matrix and distance matrix are both derived from the navigation function based on Emap API. We proposed a heuristic approach ACO+PLS to solve the minimum loss model of coldchain logistics. The experiments show that our proposed method has strong applicability and potential advantages in coldchain distribution. In future studies, a combination of multiple spatial information technologies such as geographic information technology and remote sensing technology can be realized to make the problem more practical.
Data Availability
The data used to support the findings of this study are included within the article.
Conflicts of Interest
The authors declare that they have no competing interest.
Acknowledgments
This work was supported by Scientific Research Project of Beijing Municipal Education Commission (KM201911417006).