1. Introduction
With globalization and technological advances, the frequency and impact of emergencies are increasing. For example, the New Crown epidemic in 2020, the Australian Hill fires in 2021, the Texas blackout in 2020, and the Indonesian earthquake in 2019 caused shortages or difficulties in shipping supplies. Emergency logistics systems that can respond quickly depend on large safety stocks. Large safety stocks can significantly raise the total cost of emergency logistics. Safety stock is the number of supplies increased to cope with uncertainty, which directly affects the reliability and speed of emergency response. Therefore, setting safety stock is an important consideration in the optimization problem of mobile emergency logistics.
For the multi-stage site selection problem of emergency logistics, based on the uncertainty and ambiguity of disaster relief information, Zhu [
1] proposed a cooperative optimization model with a comprehensive evaluation framework as the supplier of emergency materials. A multi-attribute grouping decision ranking method is used to select the best emergency material suppliers. Multi-objective fuzzy optimization was carried out in three emergency response phases: pre-disaster, disaster, and post-disaster. Increased demand uncertainty sometimes leads to lower levels of safety stock and increases total storage costs. In order to limit the excess risk, Zhang [
2] proposed relying more on capital stock to reduce the reliance on safety stock. Based on a continuous nonlinear formulation that integrates location, allocation and inventory decisions, Puga [
3] proposed a location-inventory siting model for large supply chains with uncertain demand. That model includes transportation cost, cycle stock, safety stock, ordering and facility activation. According to the problem of preference uncertainty in the decision-making process, Liu [
4] presents a Group Decision-Making (GDM) method with Interval Linguistic Fuzzy Preference Relationship (ILFPR). To verify the applicability and credibility of the proposed method, the article is validated with case studies. In response to the problem of high uncertainty in the evolutionary direction of emergencies, Fei [
5] offered a hybrid decision-making approach that takes into account intuitionistic fuzzy environments, linguistic environments, and their hybrid environments. The article verifies the effectiveness of the proposed model in the case of flooding in China. Wu [
6] solved the rescue path selection problem in the uncertain environment of vehicular traffic accidents by minimizing the vehicle travel time of nodes and road sections. The risk coefficient of the rescue path is calculated based on the elastic time window, and decision support is provided for the path selection of emergency rescue. To solve the emergency demand uncertainty problem, Ge [
7] introduced a two-stage model. The model is used to calculate the problem of the siting and material distribution of emergency logistics centers in the region. The NSGA-II algorithmic site selection model is combined to solve the problem. In order to solve the emergency rescue problem for driver-passenger hijacking incidents, Luo [
8] proposed a multi-vehicle mutual rescue (MMR) model for cab platforms. The Pareto evolutionary algorithm (SPEA II), Frank-Wolfe algorithm, and two-stage coding method are applied to solve the NP-hard problem of the MMR model. The MMR model is examined using road information and O-D traffic, and the utility of the hybrid algorithm and the MMR model is analyzed. In response to the problem of uncertain production disruptions affecting the supply chain during emergencies, Zhu [
9] investigated a new recovery strategy. This strategy utilizes investments to adjust the speed and duration of capacity recovery. A two-stage stochastic programming model (RTSPM) that simulates the recovery behavior in response to different disruptions to avoid risks is proposed. To justify the efficiency of the method, the article proposed a trust-region-based decomposition method to solve the RTSPM. To respond to the emergency relief problem of uncertainty in the last-mile delivery time, Zhang [
10] proposed an opportunity constraint model. This model considers emergency fair siting as emergency logistics centers (ELCs). For the contingency decision-making problem with uncertain information, Li [
11] proposed a novel consensus model to manage the non-cooperative behavior of experts in large-scale group decision-making problems. This model introduces a group consistency index that considers both the fuzzy preference value and the degree of cooperation to detect the non-cooperative behavior of experts. Among the studies related to safety stock, de Kok [
12] proposed a typology of multilevel inventory management.
Extensive research on multilevel inventory management under uncertain demand is categorized and reviewed. Maximizing the total cost of network population coverage and minimizing the risk of network management, Hasani [
13] proposed a humanitarian-based approach to the stockpile grouping problem. This model optimizes the number of inventory groups and service levels, the distribution of relief supplies, the location of relief facilities, and the distribution of relief services. Based on blood expiration, shortage and capacity, Ma [
14] developed a dynamic emergency blood collection model. It was found that the safety stock, target stock level and emergency blood demand estimation could be moderately increased. This allows for emergency blood stock control parameters. To minimize the expected total cost per unit of time, Bono [
15] proposed an inventory strategy. By considering energy costs, the article calculated the optimal size of order quantity, safety stock and inventory cycle length. In order to design dynamic multi-commodity SSC networks to compute contingency safety stocks and shared safety stocks, Zadeh [
16] proposed a mixed integer nonlinear programming (MINLP) model and mixed integer linear programming (MILP) model. To demonstrate the stochastic humanitarian inventory control model, Ozguven [
17] validated the minimum safety stock level for contingency inventory through a case study. Paterson [
18] presented two methods of centralizing inventory. Passive transshipment transfers inventory elsewhere in the network to cope with shortages at one location. Active inventory redistribution is carried out to minimize the possibility of future shortages. Aiming at the uncertainty of the actual demand at the demand point and the vehicle transportation time, Huali Sun [
19] investigated a robust optimization model for the location path of emergency facilities. The article minimizes the sum of rescue time for material delivery to the demand point.
To summarize, existing research on emergency logistics siting problems mainly focuses on emergency uncertainty, safety stock strategy, and carrying out siting problems in different contexts. This paper addresses the multilevel siting problem of mobile emergency logistics, considering safety stock under contingency. To solve the problems of frequent demand for emergencies and high emergency demand uncertainty, we calculate the emergency demand through statistical population distribution and propose a multilevel siting model for mobile emergency logistics, considering safety stock. The mobile emergency logistics system consists of logistics centers, distribution storage warehouses, and emergency distribution facilities at three levels. The logistics center can process national-level logistics, demand information and ensure that emergency supplies reach the disaster area as soon as possible based on diversified transportation modes. As the relay station of the mobile emergency logistics system, distribution storage warehouses are responsible for the temporary storage and secondary distribution of emergency supplies. The emergency distribution facility is the final link of the mobile emergency logistics and is responsible for delivering emergency supplies to the affected population. Therefore, a three-level architecture is used to construct the multi-level siting model of mobile emergency logistics. Minimize the cost of emergency logistics storage and transit under emergencies. Finally, the optimized immuno-optimization algorithm is used to solve the problem. We verify the effectiveness of the mobile emergency multilevel siting model by comparing and analyzing the traditional deployment scheme and heuristic deployment. Finally, appropriate management decision suggestions are given for the characteristics of emergencies in different scenarios.
4. Analysis of Examples
4.1. Description of the Example
A total of 148 alternative points were selected for mobile emergency logistics multilevel site selection in the Chongqing-Sichuan region. The center of mass of each administrative county and district is regarded as an alternative point in the region. The population distribution of each county and district is used to determine the demand for each alternative point. The fixed site selection cost is divided by 10 for the average house price in Sichuan and Chongqing. The normalization of alternative site coordinates and population distribution is shown in
Figure 3. The population distribution data for Chongqing and Sichuan were obtained from the seventh population census. Population density distribution data is obtained from the World Pop Hub. The coordinates of the demand points are obtained by calling the Gaode map to find the center of mass of the region through the Data Map. The fixed cost of the alternative point location is equal to that shown in the parameter setting in
Table 2. The alternative sites in Sichuan and Chongqing are summarized in
Table 3.
4.2. Algorithm Analysis
The algorithm was programmed and experimented with using MATLAB2020B. The experimental computer environment was an Intel 11th Gen Core
[email protected] quad-core with 16 G of RAM. Sensitivity analysis of IIA was performed, and it was found that the population size affects the search range of the algorithm.
Theoretically, increasing the number of iterations increases the likelihood that the algorithm will find a globally optimal solution. In our experiments, we took 50 times, that is 300 times, as the maximum number of iterations after the solution results were stabilized and there were no more changes. The crossover probability and variance probability are used to balance the local optimal solution and the global optimal solution, so the crossover probability is taken as 0.5 for stabilizing the speed and aggregation of the search process. The variance probability is taken as 0.4 to increase the population’s diversity and not make the search process too scattered and difficult to converge. The diversity evaluation parameter prevents the solutions from being too similar and falling into local optimal solutions, and 0.95 can effectively ensure the diversity of the solution results. The population size of 200 is determined mainly based on the number of candidate points. This not only helps to improve the search space of the algorithm but also ensures its performance while considering computational efficiency. The memory capacity is the maximum capacity to store the optimal solution. It is taken as 40 in order to guide the algorithm in the search process. It prevents the algorithm from falling into a local optimal solution, but it is not too large to affect the computational efficiency.
In order to test the potential solution space, a sensitivity analysis was performed on population size. Solutions were performed for population sizes of 50, 100, 150, 200 and 250 without changing other parameter settings. To address the effect of population size on the value of the objective function, an improved immuno-optimization algorithm was tested. The results show that the population size affects the initial solution level directly, and the larger the population size, the better the initial performance. As the number of iterations increases, population size is negatively correlated with affinity. The larger the population, the more stable the convergence is. However, the number of iterations required to stabilize the relationship increases. After many iterations, the population size does not have much effect on the affinity value. This is because both the maximum and minimum population size curves eventually converge to a closer affinity value. Therefore, it is reasonable to set the population size to 200, as shown in
Figure 4, with Xlabel representing the number of iterations.
The unimproved immune optimization algorithm (IA) and the improved immune optimization algorithm (IIA) were selected for performance comparison. The parameter settings of the enhanced algorithm are shown in
Table 4. The solution comparison with the Genetic Algorithm (GA) and Particle Swarm Algorithm (PSO) using both algorithms according to the same parameters, respectively, is shown in
Figure 5. To verify the effectiveness of the improved immune optimization algorithm, its adaptation values were compared with those of the other three basic intelligent optimization algorithms.
From
Figure 6, we can see that the adaptation value of IIA is higher than that of IA, GA, and PSO algorithms in the early iteration. As the number of iterations increases, the local search ability and global search ability of the IIA algorithm are further enhanced. The adaptation value of IIA algorithm is better than that of the PSO and IA algorithms. PSO and IA are more stable throughout the iterations, and the final search result is slightly better than GA’s. Still, PSO tends to fall into the local optimum at the beginning of the iteration. In addition, the GA optimum is more stable at the beginning of the algebraic iteration. Still, with the increase in iterations, the GA falls into the local optimum solution, while the IIA solves this problem. It jumps out of the local optimum, enhances the local search ability, and improves the data quality of IA.
IIA outperforms most classical intelligent algorithms by combining adaptation value and stability. It effectively reduces the cost of emergency logistics, has a robust global search capability, and is not easily trapped in the local optimum. Therefore, the IIA solver obtains the siting result of the siting model. The solution can provide a certain reference for the actual emergency logistics site selection planning.
4.3. Result Analysis
The coordinates of mobile emergency facilities in 148 counties and districts were selected in Sichuan and Chongqing. The emergency logistics centers were chosen using the improved immune optimization algorithm. As shown in
Figure 6, two logistics centers were deployed in Jinniu District, Chengdu and Hechuan District, Chongqing. Distribution warehouses were selected among the alternative sites in the remaining districts and communities. Materials were transported to the distribution warehouses through the logistics centers.
The relationship between the affinity value, solution time, and the number of warehouses is shown in
Figure 7.
An increase in the number of warehouses will increase the level of mobile emergency logistics. Affinity decreases as the number of warehouses increases. Still, the affinity curve decreases with the increase in the number of warehouses, slows down significantly, and levels off around the number of warehouses equal to 18. Due to the intelligent optimization algorithm, the solution time becomes longer as the number of allocated storage points increases. The solution time also fluctuates occasionally as the number of allocated storage bins increases. Generally speaking, the higher the safety stock coefficient of the distribution warehouse, the larger the safety stock is. The larger the safety stock, the lower the transit cost. The statistics of the number of emergencies at all levels in the last 20 years in all districts and counties of Sichuan and Chongqing are shown in
Table 5 and plotted as shown in
Figure 8.
Level 1 to 4 emergencies have approximately the same probability of occurring in each region. Each region’s safety stock factor assignment is based on a summary of the number of emergencies. The safety stock coefficients were assigned to four intervals (0.2, 0.3, 0.4, 0.5) for the total number of emergencies in the last 20 years (0–16, 17–30, 31–45, 46–70). The number of warehouses is equal to 18. Warehouses in each region are assigned safety stock coefficients according to
Table 4 for site selection. The results of the warehouse siting are shown in
Figure 9. In summary, further analysis leads to the following:
Based on the above site selection results, the relationship between inventory, transshipment, and total costs is measured by adjusting the safety factor, as shown in
Figure 10.
Inventory costs include contingency and safety inventory costs, and transshipment costs include horizontal and vertical transportation costs. In general, inventory costs become higher as the safety stock factor increases. Transit costs decrease as the safety stock factor increases. The total cost of emergency logistics is mainly affected by transshipment costs. Total emergency logistics costs increase with the rise in the safety stock factor. In addition, the calculations based on the above overview also lead to:
(1) Safety stock and transshipment costs are negatively correlated, i.e., an increase in safety stock will reduce transshipment costs. The ratio of safety stock and trans-shipment fee unit costs affects the setting of the safety stock strategy. The larger the proportion of safety stock costs, the greater the impact on total costs. The smaller the transit fee balance, the greater the impact on total costs. In general, transshipment costs mainly affect the total cost of emergency logistics. Therefore, a higher safety stock factor in emergency-prone areas can reduce the transshipment cost more effectively, and it is reasonable and feasible to assign a value to the safety stock factor according to the statistical level of emergency events.
(2) The deployment of the Emergency Logistics Center has been effective. Contingency logistics centers can only be deployed in smaller quantities because of higher deployment costs and inventory levels. Transit costs mainly affect the total cost of emergency logistics. Emergency logistics centers cover several warehouses. The more the logistics transfer effect plays, the stronger it is, and warehouses can still move the transfer. Therefore, from the perspective of the mobile emergency logistics system, the actual coverage of the emergency logistics center has been significantly improved under the safety stock strategy.
(3) There is a lower limit on the cost of mobile emergency logistics. The total cost of emergency logistics is limited by population distribution, transshipment costs, and inventory costs and cannot be further reduced after it reaches a certain level. The removal of restrictions on the number of emergency logistics centers and warehouses to be deployed has resulted in an overall leveling off of the total cost of emergency logistics. The total cost of contingency logistics will not decrease as a result of an increase in the number of contingency logistics centers and warehouses. Non-core counties and districts with sparse population distribution, remote distance, and low safety stock coefficients covering their emergency needs mainly rely on the neighboring warehouses for transshipment. This suggests that adopting a hierarchical coverage of mobile emergency logistics is a more effective strategy in emergencies.
In summary, we give the following management suggestions: 1. In areas where emergencies occur more frequently, setting up a higher level of safety stock can effectively reduce the total cost of emergency logistics; 2. Emergency logistics centers should be built in areas where the population is more concentrated so that they can more effectively respond to the emergency needs of many people during emergencies; 3. In areas where the population distribution is more sporadic, the number of distribution warehouses should be increased and their coverage expanded to meet emergency needs more effectively.