1. Introduction
With the development of China’s urban logistics industry, the use of fossil energy increases year by year, which leads to environmental pollution. Electric vehicle distribution has the advantage of “low energy consumption, low pollution.” “Low energy consumption” refers to minimal use of energy resources, while “low pollution” indicates reduced emission of harmful substances. Together, they describe environmentally-friendly practices that minimize energy use and environmental impact, crucial for sustainable development and ecological preservation. For this reason, some logistics enterprises began to use electric vehicle distribution. However, electric vehicles are limited by their short driving range and susceptibility to urban time-varying road networks. During large-scale distribution processes, drivers often experience “range anxiety.” Range anxiety refers to the psychological stress experienced by electric vehicle (EV) drivers due to concerns about insufficient battery range to complete a trip or reach the next charging station. This phenomenon is particularly relevant in urban logistics, where frequent stops and varying traffic conditions can exacerbate range uncertainty. This anxiety results in the underutilization of battery capacity. In addition, with the diversification of people’s demand for commodities, distribution centers need to send out vehicles of different sizes according to the types of products demanded, the amount of customer demand and other factors, which further increases the difficulty of optimizing the distribution path of electric vehicles. Therefore, in the process of distribution services, rational planning of vehicle paths and distribution modes, as well as the adoption of appropriately configured fleets and charging strategies, have become urgent problems to be solved.
The electric vehicle routing problem (EVRP) is an extension of the classical vehicle routing problem (VRP) and focuses on optimizing delivery routes for electric vehicles (EVs) while considering their unique characteristics, such as limited battery capacity, energy consumption rates, charging station locations, and charging time requirements. The primary objective is to minimize total delivery costs while ensuring that vehicles can complete their routes without running out of charge, often incorporating constraints related to charging infrastructure, vehicle range, and time windows for customer deliveries. Numerous scholars have conducted research on various variants of this problem. Schneider et al. [
1] studied the electric vehicle routing problem with time windows and charging stations. Ahmed et al. [
2] investigated the vehicle routing problem with time windows and proposed a hybrid ant colony optimization algorithm for its resolution. Jiang et al. [
3] addressed the issue that charging navigation cannot simultaneously optimize the charging destination and route planning for electric vehicles. They established a bi-level stochastic optimization model for electric vehicle charging navigation that considers multiple uncertainty factors and proposed a hierarchical enhanced deep Q network-based method to solve the aforementioned stochastic optimization model. He et al. [
4] considered the electric vehicle routing problem with multi-temperature co-distribution and proposed a mathematical model with soft time windows. They improved the ant colony optimization algorithm by incorporating the 2-opt algorithm, and experimental results showed that this method effectively reduces delivery costs and improves delivery efficiency. Veena et al. [
5] focused on charging prices and proposed a time-of-use pricing-based efficient electric vehicle routing optimization method using the bat algorithm, aiming to reduce electricity costs and total travel distance.
With the gradual advancement of research, numerous scholars have begun to explore more variants of the electric vehicle routing problem (EVRP) to address urban logistics distribution challenges. For instance, some researchers have focused on heterogeneous fleet distribution strategies. The heterogeneous fleet distribution strategy involves the use of a mixed fleet of vehicles with varying capacities, energy source, and operational characteristics. This strategy enables logistics enterprises to tailor vehicle assignments based on specific delivery requirements, thereby improving efficiency, reducing costs, and supporting sustainable practices. Wang et al. [
6] constructed an integer programming model considering mixed fleets to solve the “last mile” delivery issue in rural logistics. Afsane et al. [
7] established a mixed-integer programming model for route selection involving conventional and hybrid fleets in urban logistics distribution and employed an adaptive large neighborhood search (ALNS) algorithm for solution. Wang et al. [
8] investigated the heterogeneous fleet vehicle routing problem for hazardous materials and solved it using an adaptive fuzzy large neighborhood search algorithm. Sarbijan et al. [
9] developed a multi-type vehicle routing optimization model under mid-route replenishment scenarios and utilized a particle swarm optimization algorithm combined with simulated annealing for solving. Sun et al. [
10] proposed a multi-vehicle collaborative segmented transshipment model to improve the efficiency of fresh product distribution, employing the k-means clustering algorithm to determine transshipment points and an adaptive multi-objective ant colony optimization algorithm for solution. The aforementioned studies predominantly adopted single-center distribution models or multi-center closed models. To address large-scale customer distribution challenges, some scholars have shifted their focus to multi-center semi-open vehicle routing problems. The semi-open distribution model refers to a logistics strategy where vehicles are allowed to return to the nearest distribution center after completing their delivery tasks, rather than being restricted to their original departure point. This approach reduces travel distances, optimizes resource utilization, and enhances operational flexibility, particularly in large-scale distribution networks. For such problems, Nistha et al. [
11] investigated a multi-center vehicle routing problem for surplus food redistribution and solved it using an elite genetic algorithm. Alaia et al. [
12] treated the multi-vehicle, multi-depot, semi-open vehicle routing problem with time windows as a multiple-criteria optimization problem and applied a genetic algorithm with elite selection strategies. Nunes et al. [
13] resolved the multi-depot semi-open vehicle routing problem with time windows using intelligent general variable neighborhood search and general variable neighborhood search algorithms. Li et al. [
14] demonstrated that shared warehouses yield better economic benefits compared to non-shared ones in their study on multi-depot vehicle routing problems. Ruiz et al. [
15] developed a multi-depot semi-open fuel vehicle routing optimization model with capacity and distance constraints, employing an adaptive large neighborhood search algorithm for solution.
In addition, EV urban distribution has strong timeliness, so more scholars have studied the time-dependent EV vehicle path problem. Ichoua et al. [
16] introduced the “first in, first out” criterion to the time-dependent vehicle path problem for the first time, and represented the dynamic road network through the stage speed time-dependent function. Gmira et al. [
17] considered the variation in vehicle travel time and path length under a time-varying road network and used the taboo search algorithm to solve it. Ehsan et al. [
18] designed a variable neighborhood taboo search algorithm to solve the time-dependent vehicle path problem, which employs a granular local search mechanism in the reinforcement phase and a forbidden oscillatory mechanism in the diversification phase of the variable neighborhood search. Stoia et al. [
19] studied the time-dependent crowdsourcing problem in a time-varying network, as well as the time-dependent crowdsourced delivery path optimization problem, which was solved using an adaptive large neighborhood search algorithm.
The aforementioned research achievements have further enhanced the quality and efficiency of electric vehicle (EV) distribution in urban logistics. However, few scholars have considered the impact of different charging strategies on key metrics such as total costs and energy consumption. To address this gap, researchers have begun to investigate EV charging strategies, which involve formulating various charging plans and methods based on battery status, charging infrastructure distribution, and delivery task requirements, aiming to optimize distribution efficiency, reduce costs, and extend battery lifespan. Desaulniers et al. [
20] compared full charging strategies with partial charging strategies and employed a branch-price-and-cut algorithm to compute optimal solutions for variants of the EVRP with time windows (EVRPTW) model. Meng et al. [
21] considered battery swapping strategies, established an EVRP model with soft time window constraints to minimize total costs, and solved it using an ant colony optimization algorithm. Wang et al. [
22] developed a multi-objective optimization model for hybrid EV routing and charging strategies. Lian et al. [
23] and Alejandro et al. [
24] studied EV routing problems based on nonlinear charging functions and compared them with linear charging functions, demonstrating that ignoring nonlinear charging leads to excessively high solution costs. Yuan et al. [
25] introduced different charging strategies for non-emergency patient transport routing optimization, constructing a multi-objective model to maximize patient satisfaction and minimize transport costs. Liu et al. [
26] proposed a mathematical model for shared EV infrastructure planning and smart charging strategies utilizing spatiotemporal behavior and data extracted from shared EV trajectory datasets to quantify infrastructure demand. Zhang et al. [
27] addressed the issue of long EV charging times by proposing a time-dependent hybrid charging strategy that combines full charging, partial charging, and battery swapping strategies to reduce the impact of charging time on routing decisions.
By reviewing the aforementioned literature, the following research gaps are identified.
1. With the diversification of customer demands, logistics enterprises may dispatch different types of vehicles from multiple distribution centers based on product characteristics. However, research that simultaneously considers multi-center heterogeneous fleets is relatively scarce. For instance, Afsane et al. [
7] fails to account for the multi-center semi-open distribution model, while Ruiz et al. [
15] considers the multi-center semi-open distribution model, but overlooks the heterogeneity of the fleet.
2. The consideration of charging strategies is often limited. For example, Meng et al. [
21] only examines battery swapping strategies, and although Desaulniers et al. [
20] and Zhang et al. [
27] explore multiple charging strategies, they do not analyze the variations in multiple performance metrics.
3. The electric vehicle routing problem (EVRP) is an NP-hard problem with high solution space complexity. Traditional metaheuristic algorithms employed in some studies, such as the ant colony algorithm, simulated annealing algorithm, and particle swarm optimization algorithm used in Sarbijan et al. [
9] and Meng et al. [
21], often struggle to ensure satisfactory solution quality and computational efficiency, as they are prone to converging to local optima.
Based on the existing research, this paper investigates the following issues.
1. Adopting a multi-depot semi-open distribution mode and a heterogeneous fleet strategy, by adjusting the number of vehicles of each type and coordinating distribution plans across multiple depots, to achieve the goals of reducing distribution route length and minimizing total distribution costs;
2. Constructing a nonlinear discharge function by comprehensively considering actual vehicle load, driving speed, vehicle parameters, and road impedance to make the proposed model more aligned with real-world scenarios;
3. Designing an improved adaptive large neighborhood search algorithm to solve the proposed problem based on its characteristics;
4. Comparing the economic benefits and distribution efficiency brought by different charging strategies and providing scientific management insights for logistics enterprises, which helps reduce operational costs, improve distribution efficiency, and support decision-making optimization under diverse demands.
2. Problem Description and Assumptions
Logistics distribution primarily refers to the activities undertaken by logistics companies to ensure the normal operation of product supply, which involves obtaining raw materials from warehouses within the logistics enterprise or distributing products to subordinate suppliers. To effectively arrange distribution tasks, companies typically notify distribution centers in advance of the demand for various goods, and the distribution centers organize the delivery operations based on the actual demand for goods and the distribution of customer nodes. For large-scale distribution tasks with relatively dispersed customer points, employing a single distribution center or a multi-distribution center closed delivery model may lead to long delivery routes and difficulties in vehicle scheduling, making it challenging to meet customer-defined time windows. In contrast, the multi-center semi-open delivery model allows vehicles to return to the nearest distribution center after completing deliveries, thereby reducing delivery distances and total delivery costs.
In terms of model construction, this paper formulates an objective function aimed at minimizing the sum of fixed costs, travel costs, energy consumption costs, and time window penalty costs utilizing a nonlinear discharge function to achieve the minimization of total delivery costs. The proposed model encompasses multiple distribution centers and multiple customer points, with the coordinates of all distribution centers and customer points being known. All distribution centers are equipped with vehicles of different load capacities (Type I and Type II: Type I vehicles have larger loads and energy consumption, while Type II vehicles are smaller). Both types of vehicles can depart from any distribution center, and upon completing their delivery tasks, can return to any nearby distribution center. Additionally, during the delivery process, road congestion may occur; therefore, the multi-center multi-vehicle electric vehicle routing problem constructed in this paper incorporates the influence of a time-varying road network, where vehicle speeds vary across different time periods.
Based on the aforementioned problem description and assumptions, this paper makes the following assumptions:
1. The locations of each distribution center, customer point, and charging station are known, as are the demand quantities and time windows for each customer point.
2. The demand of each customer point cannot be split, meaning that each customer point is served exclusively by one vehicle. However, each electric vehicle (EV) is capable of serving multiple customer points.
3. Vehicles are categorized into Type I and Type II, with differing parameters between the two types.
4. Electric vehicles start in a fully charged state, and if the vehicle’s battery level drops below a threshold (12%) during transit, the driver experiences “range anxiety” and must immediately cease the delivery task to head to the nearest charging station.
5. Vehicles can recharge multiple times en route, with sufficient charging stations available and identical charging rates across different charging stations, allowing multiple vehicles to charge simultaneously without waiting in line.
6. During service, vehicles turn off their engines and do not consume energy.
7. The actual load of the electric vehicle must not exceed its maximum capacity.
8. If the vehicle fails to deliver within the customer’s acceptable time window, a penalty cost is incurred.
Based on the above problem description and assumptions, the type of urban logistics distribution model proposed in this paper is depicted in
Figure 1.
4. Algorithm Design
This paper proposes a multi-center and multi-type electric vehicle routing problem with a time-varying road network as an extension of the electric vehicle routing problem. The adaptive large neighborhood search (ALNS) algorithm, introduced by Ropke and Pisinger [
31] in 2006, has demonstrated superior performance in solving vehicle routing problems (VRPs) and has been successfully applied to various VRP variants. However, the ALNS algorithm has certain limitations. During the search process, it may converge to local optima, and its solution efficiency and quality can be compromised if the operators are inefficient. To address these limitations, an improved adaptive large neighborhood search (IALNS) algorithm is designed to solve the multi-center time-dependent electric vehicle routing problem with time windows (MCTEVRPTW).
In the IALNS algorithm, more efficient operators are designed to expand the solution search space. Based on the problem characteristics, damage and repair operators specific to vehicle paths are constructed. An adaptive strategy is introduced to select efficient operators dynamically, and a simulated annealing-based solution acceptance criterion is incorporated to allow for the acceptance of inferior solutions within a certain probability, thereby avoiding convergence to local optima. The fundamental framework of the algorithm and its improvement scheme are outlined as follows.
4.1. Basic Components of ALNS
4.1.1. Probability of Accepting a New Solution Based on Simulated Annealing Algorithm
During the iterative optimization search process, if the newly generated solution outperforms the current solution, it is adopted immediately. Otherwise, the new solution is accepted with a certain probability. The acceptance probability formulae for the new solution proposed in this study are presented in Equations (35) and (36):
In Equations (35) and (36), denotes the value of the objective function of the new solution; denotes the value of the objective function of the current solution; and T denotes the value of the current temperature. The formula for the temperature is: , where a is the cooling rate of the simulated annealing temperature. The initial value of a is set to , where is the rate, assuming that , i.e., in the initial stage, there is a probability of 0.5 to be accepted when the new solution obtained is inferior to the initial solution. After iterations of the algorithm, the temperature will decrease by multiplying it with a constant simulated annealing cooling rate a, and thus the probability of accepting a new solution inferior to the initial one decreases to ensure the convergence of the solution.
4.1.2. Adaptive Strategy
An adaptive strategy is introduced to ensure that the weights of each operator are updated to the optimal state in each iteration loop. In the IALNS algorithm, the weights of each destruction and repair operator are denoted
. Higher weights indicate better performance in the last iteration and an increased probability of being selected in the next iteration. In each iteration, the weights are updated according to
. The weight update formula is shown in Equation (37):
where
is the weight adjustment speed factor and
,
is the score obtained by operator
i in the previous iteration, and
is the number of times operator
i was used in the previous iteration. The operator weights are then normalized on the basis of classification. During the iteration process, the selection of destruction and repair operators follows the roulette rule.
4.1.3. Algorithm Scoring Strategies
In terms of adaptive large neighborhood search algorithms, there is a lack of research on improving the scoring strategy. Traditional studies generally adopt fixed values, such as = 30, = 20, = 10, etc. In this paper, we propose a randomized scoring mechanism, in which , , and are set as random numbers in the intervals of [27, 33], [17, 23], and [7, 13], respectively. This randomized scoring mechanism aims to enhance the global search capability to increase the probability of the algorithm jumping out of the local optimal solution.
4.2. SOM Initialization
The k-means clustering algorithm is employed to initially group the nodes, and the self-organizing mapping (SOM) method is utilized to initialize the paths of the electric vehicles (EVs). This approach provides a reasonable initial solution for the proposed model, enabling the subsequent optimization algorithm to converge more efficiently toward the global optimal solution. The execution algorithm is outlined in Algorithm 1, with the following steps.
Step 1: At the beginning of the function, declare the global variable P, which contains parameters and data relevant to path planning, such as node coordinates, time windows, and the number of routes.
Step 2: Apply the k-means clustering algorithm to group the node coordinates and output the cluster labels and cluster centers for each node.
Step 3: Traverse each cluster and extract the node indices and time window information associated with that cluster. Sort the nodes according to their time windows, construct a new path, and append the end node to the path.
Step 4: Calculate the path information for each route, including distance, time, and cost metrics, and store the results in a pending collection.
Step 5: Refine and optimize the generated path information to ensure the rationality and feasibility of the paths, thereby providing effective initial solutions for subsequent optimization algorithms.
Algorithm 1: SOM initialization |
function kmeans_initial () global P
[val, C] = kmeans (P.cor (2:P.node_N (1) +P. node_N (2),:), P. line_N, “Replicates,” 10) task = []
for n from 1 to P. line_N do rn = [1; find (val == n) + 1] tl = P.t_window (rn, 2) tl (1) = 0; tl(end) = 0 [~, ind] = sort(tl) route = [rn(ind)’ − 1, 0]
if n == 1 then task = cal_one(route) else task(n) = cal_one(route) end if end for
task = Amend(task) end function |
4.3. Destruction Operator Design
4.3.1. Greedy Destruction Operator
The greedy destruction operator evaluates the impact of node removal on path performance by disrupting the existing path structure, thereby selecting the optimal deletion operation. The execution algorithm is detailed in Algorithm 2, with the following steps.
Step 1: Extract the input task information and record the current path along with its associated metrics.
Step 2: Randomly determine the number of nodes to be removed, ranging from 1 to 5, to introduce randomness and uncertainty into the process.
Step 3: Iterate through each route and each removable node, calculate the path performance after node removal, and record the resulting reduction in path performance.
Step 4: Identify the node whose removal results in the largest reduction in path performance, delete it, update the current route’s path information, and return the modified task information.
Algorithm 2: Greedy destruction operator |
function destroy_greedy(task) global P
task = task. task task1 = task linen = length(task)
destroy_n = random_integer (1, 5) node = []
for nn from 1 to destroy_n do df = []
for n from 1 to linen do for j from 2 to task1(n). len − 1 do route_n = task1(n). route route_n(j) = remove(route_n(j))
ret = cal_one(route_n) dfj = task1(n). fit − ret.fit
df. append ([n, j, dfj]) end for
end for [~, ind] = max (df [:, end]) n = df (ind, 1) j = df (ind, 2)
route_n = task1(n). route node. append(route_n(j)) route_n(j) = remove(route_n(j))
task1(n) = cal_one(route_n) end for return task1 end function |
4.3.2. Random Destruction Operator
The random destruction operator optimizes path planning for electric vehicles (EVs) by randomly removing nodes. The core concept of this operator is to assess the impact of node removal on path performance by randomly selecting and deleting nodes from existing paths. The execution algorithm is outlined in Algorithm 3, with the following steps.
Step 1: Extract the input task information and record the current path along with its associated performance metrics.
Step 2: Randomly determine the number of nodes to remove, denoted destroy_n, with a range of 1 to 5.
Step 3: Randomly select a route that contains a sufficient number of demand points, and then randomly choose a valid node position kk for the removal operation.
Step 4: Record the node to be deleted, remove it from the current route’s path, and calculate the performance of the updated path.
Algorithm 3: Random destruction operator |
function destroy_random(task) global P
task = task. task task1 = task lineN = length(task)
destroy_n = random_integer (1, 5) node = []
for nn from 1 to destroy_n do while true do n = random_permute (linen, 1) route = task1(n). route if task1(n). len > 2 then break end if end while
k = random_permute(task1(n). len − 2, 1) + 1
node. append(route(k))
route(k) = remove(route(k)) task1(n) = cal_one(route) end for
return task1 end function |
4.3.3. Vehicle Path Destruction Operator
The vehicle path destruction operator optimizes the electric vehicle (EV) path by randomly removing all intermediate nodes within a selected path. This operator evaluates the impact of node removal on path performance by disrupting the existing path structure. The execution algorithm is detailed in Algorithm 4, with the following steps.
Step 1: Extract the input task information and record the current path along with its associated performance metrics.
Step 2: Randomly select a route n and verify its length len to ensure that it contains at least two nodes (start and end).
Step 3: After identifying a suitable route, record its intermediate nodes, remove these nodes from the current route’s path, calculate the performance of the updated path, and update the current route information.
Step 4: Upon completing the destruction and updating of the path, return the modified task information.
Algorithm 4: Vehicle path destruction operator |
function destroy_route(task) global P
task = task. task task1 = task linen = length(task)
while true do n = random_permute (linen, 1) if task(n). len < 2 then continue end if
route = task(n). route node = route (2: end-1) route (2: end-1) = remove (route (2: end-1)) task1(n) = cal_one(route)
break end while
return task1 end function |
4.4. Repair Operator Design
4.4.1. Greedy Repair Operator
The greedy repair operator selects the insertion position that minimizes the impact on overall path performance by evaluating the effect of inserting each node on the path’s performance. The execution algorithm is outlined in Algorithm 5, with the following steps.
Step 1: Extract the input s_destroy information to record the current path and the removed nodes.
Step 2: Initialize an empty array to store the performance changes of each route at different insertion positions.
Step 3: Iterate through each route and each potential insertion location, calculate the path performance after inserting the nodes, and record the results.
Step 4: Identify the insertion location that results in the smallest increase in path performance, perform the insertion, and update the current route information.
Algorithm 5: Greedy repair operator |
function repair_greedy(s_destroy) global P
task = s_destroy. task linen = length(task)
for node in s_destroy. node do df = [] for n from 1 to linen do for j from 1 to task(n). len − 1 do route_n = insert(task(n). route, j, node) ret = cal_one(route_n) dfj = ret.fit − task(n). fit
df. append ([n, j, dfj]) end for end for
[~, ind] = min (df[:, end]) n = df (ind, 1 j = df (ind, 2)
route_n = insert(task(n). route, j, node) task(n) = cal_one(route_n) end for
return task end function |
4.4.2. Random Repair Operator
The stochastic repair operator selects the insertion position that minimizes the impact on the overall path performance by evaluating the impact of the insertion of each node on the path performance or randomly selects the insertion position if the path performance is degraded. Its execution algorithm is shown in Algorithm 6, with the following steps.
Step 1: Extract information from the input s_destroy to record the current path state and deleted nodes.
Step 2: Create empty array df to store the performance change information of each line at different insertion locations.
Step 3: Iterate over each line and its possible insertion locations, and calculate and record the path performance after inserting nodes.
Step 4: Select the insertion location based on the performance variation. If there is a decrease in performance, use the roulette selection method to randomly select the insertion location. Conversely, select the location with the smallest increase in performance for node insertion and update the line information.
Algorithm 6: Random repair operator |
function repair_random(s_destroy) global P
task = s_destroy. task linen = length(task)
for node in s_destroy. node do df = []
for n from 1 to linen do for j from 1 to task(n). len − 1 do route_n = insert(task(n). route, j, node) ret = cal_one(route_n) dfj = ret.fit − task(n). fit
df. append ([n, j, dfj]) end for end for
df1 = df (df[:, 3] < 0,:) if not empty(df1) then val = -df1[:, 3] ind = roulette_choose(val) n = df1(ind, 1) j = df1(ind, 2) else [~, ind] = min (df[:, 3]) n = df (ind, 1 j = df (ind, 2) end if
route_n = insert(task(n). route, j, node) task(n) = cal_one(route_n) end for
return task end function |
4.4.3. Vehicle Path Repair Operator
The vehicle path repair operator optimizes the EV delivery path by locally adjusting the existing path. Its execution algorithm is shown in Algorithm 7, with the following steps.
Step 1: Extract information from the current path solution and set parameters for repair.
Step 2: Randomly select the nodes or path segments to be repaired.
Step 3: Operate on the selected nodes or path segments to generate multiple neighborhood solutions.
Step 4: Evaluate each generated neighborhood solution, calculate its performance index, and select the neighborhood solution with the best performance as the new current solution.
Step 5: Update the selected neighborhood solution to the current path and record the information during the repair process.
Algorithm 7: Vehicle path repair operator |
function repair_route(current_solution) global P
for iteration from 1 to max_iterations do node_to_repair = random_select_node(current_solution)
neighborhood_solutions = [] for operation in [insert, delete, swap] do new_solution = apply_operation (current_solution, node_to_repair, operation) neighborhood_solutions. append(new_solution) end for
best_neighbor = null best_improvement = ∞
for solution in neighborhood_solutions do improvement = evaluate(solution)—evaluate(current_solution) if improvement < best_improvement then best_improvement = improvement best_neighbor = solution end if end for
if best_neighbor is not null then current_solution = best_neighbor if evaluate(current_solution) < evaluate(best_solution) then best_solution = current_solution end if end if end for
return best_solution end function |
7. Conclusions
In this paper, for the problem of time-dependent multi-distribution center heterogeneous fleet path optimization and charging strategy selection, an improved adaptive large neighborhood search algorithm is used to solve the problem. Simulation experiments are conducted using standard test cases and real cases, and the following conclusions are drawn.
1. In terms of algorithm design, the improved adaptive large neighborhood search algorithm proposed in this paper uses a new operator scoring strategy and a new solution acceptance criterion. According to the actual problem, efficient vehicle path destruction and repair operators are designed to expand the search space of the solution, effectively preventing the algorithm from falling into local optima and further improving the solution efficiency and solution quality. In addition, an adaptive strategy is introduced to ensure that the weights of each operator are updated to the optimal state in each iteration loop.
2. In terms of model construction, this study introduces several key innovations. First, a nonlinear power consumption measurement model is proposed, which comprehensively considers the effects of actual vehicle load, travel speed, vehicle parameters, and road impedance on energy consumption. This model enables a more accurate calculation of the vehicle’s power consumption during distribution services. Second, the implementation of a partial charging strategy effectively reduces the total distribution time, energy consumption, and overall distribution costs. Third, compared to traditional homogeneous fleets, the adoption of a heterogeneous fleet not only lowers the total distribution costs but also achieves an effective balance between economic and environmental benefits. Finally, in contrast to other distribution models, the multi-center semi-open distribution model significantly shortens the distribution distance and time, reduces electric energy consumption, and provides a practical foundation for the sustainable development of green logistics.
3. In terms of managerial implications for logistics enterprises, this study demonstrates significant value. First, by optimizing distribution routes and charging strategies, it substantially reduces operational costs. Through the rational scheduling of electric vehicles (EVs) and fuel-powered vehicles, vehicle load rates are increased, thereby lowering electricity and maintenance expenses. Second, the study enhances distribution efficiency and service quality. Time-dependent route optimization enables real-time responses to traffic conditions and customer demand fluctuations, ensuring timely completion of delivery tasks and improving customer satisfaction. Additionally, the flexible scheduling of heterogeneous fleets and the semi-open distribution model expand the scope of distribution, strengthening the enterprise’s market competitiveness.
4. From the perspective of environmental sustainability, the integration of EVs into logistics distribution significantly reduces carbon emissions and energy consumption. This allows logistics enterprises to better fulfill their social responsibilities and enhance their brand image. Furthermore, the optimization of charging strategies, such as the partial charging strategy proposed in this study, extends battery lifespan and reduces battery replacement costs, thereby minimizing resource waste. Finally, this research provides logistics enterprises with scientific management tools and decision-making support. By incorporating advanced optimization algorithms and data analysis techniques, enterprises can achieve intelligent management of the distribution process, improve resource utilization efficiency, and lay the foundation for future business expansion and technological upgrades.
In summary, this paper provides a feasible solution for the time-dependent electric vehicle routing problem with a multi-depot semi-open heterogeneous fleet, demonstrating significant adaptability and robustness, particularly in addressing heterogeneous fleets and distribution modes. However, this study has certain limitations. Future research could introduce more complex real-world scenarios and diverse uncertainty factors to further expand the model’s applicability and enhance its effectiveness in various logistics environments. For example:
1. The current model includes only two types of vehicles with different capacities/sizes, and future studies could incorporate a wider variety of vehicle types;
2. The waiting time for electric vehicles at charging stations due to queuing is not considered;
3. Uncertainties in customer demand and demand time windows are not addressed.
Such improvements would further enhance the adaptability of the electric vehicle routing problem and increase the practical value of the model in dynamic logistics environments.