2. Literature Review
The method of mixed operation on double routings was first used in public bus services. As early as 1987, coordination modes to find the schedule offset between different routings were used to balance loads and minimize overall cost, where the primary objective was to minimize fleet size and the secondary objective was to minimize waiting time for a given fleet size. It was shown that even when overall capacity exceeds volume on every link, there may still be one or more patterns of trip schedule that are not systematically overcrowded [
3]. Cristián et al. [
4] found that the mixed routings strategy can be justified in many cases with uneven load patterns. Since then, an extensive variety of models for mixed operation on double routings have been proposed by researchers. The researches mainly focused on operation cost and the main purpose was either to minimize the total cost or maximize the total benefit. Ji et al. [
5] studied an optimization model on bus routes to reduce the total cost. The model optimized the schedule coordination of a long routing service mode and a short-turning service mode and balanced the bus load along the route and between the two service patterns.
The relationship between operation cost and energy consumption is inseparable. With the widespread concern surrounding energy consumption, the issues of how to reduce energy consumption and capacity waste have become important for researchers. Yang et al. [
6] developed a cooperative scheduling approach to optimize the timetable so as to improve the utilization of recovery energy. Yu et al. [
7] established a two-stage model of a high-speed rail train operation plan. In the first stage, the proposed model aimed for the minimum total cost of train operation, which generated fewer train candidate sets, thereby improving the efficiency of the solution. In the second stage, the model gave an accurate flow distribution model to obtain the optimal economic efficiency and completed the design of the high-speed express train operation plan.
In addition to energy consumption, service level is also important for operators. Szeto and Wu [
8] proposed a model to reduce the number of transfers and the total travel time of passengers, thereby improving upon existing bus services. An integrated solution method was achieved to simultaneously determine the departure frequency and operating route. Cadarso and Marín [
9] developed an integrated planning model to readjust departure frequencies by recalculating the railway timetable and rolling stock. The purpose of the model was to cater for increasing passenger demand and to better adjust these flows to passenger demand, thereby alleviating traffic congestion around urban and suburban areas. Canca et al. [
10] described a short-turning strategy to optimize train operation; the optimization objective was to diminish the passenger waiting time and preserve a certain level of service quality. The model determined the service offset and turn-back stations. Huang et al. [
11] constructed a load balancing approach to optimize metro train operation. The model was based on an improved route which generalized the time utility function, considering the penalties of both in-vehicle congestion and transfers, thereby improving the service quality and reducing safety risks in metro systems, at the same time.
Many researchers considered reducing energy waste and improving service quality, at the same time. Sun et al. [
12] put forward a multi-objective optimization model by studying specific methods on train routing to optimize high-speed railway network operation. The model considered energy consumption and trains’ average travel time and also took user satisfaction into account. Yang et al. [
13] formulated a two-objective integer programming model to decrease the passenger waiting time and increase the utilization of regenerative energy, simultaneously. The model optimized the train timetable by controlling the train’s headway time and dwell time. D’Acierno and Botte [
14] developed an analytical approach to deciding driving strategies. The framework could be used for properly supporting the implementation of eco-driving strategies, from a passenger-oriented perspective. The numerical examples showed that the proposed approach is useful in determining the optimal compromise between travel time increase and energy decrease. Wang et al. [
15] investigated a multi-objective vehicle-routing optimization problem, where the objectives aimed to minimize total energy consumption and customer dissatisfaction, simultaneously. The study found a better balance between energy consumption diminishment and customer satisfaction and, finally, developed a mixed integer programming model to address the problem. Sun et al. [
16] developed a multi-objective timetable optimization model to minimize total passenger waiting time and energy consumption. The model was verified by real-world smart-card automated fare collection data and it was shown that the developed model could improve passenger service and reduce pure energy consumption more efficiently in comparison with the timetable used currently.
Many researchers considered reducing operation cost and improving service quality, at the same time. Site and Filippi [
17] chose the net benefit composed of user waiting times minus operator costs as the objective function and developed an optimization framework to obtain the maximum net benefit. The framework gave thought to variable vehicle size and service patterns over different operation periods so as to obtain an intermediate-level planning of bus operations. Chang et al. [
18] applied mixed operation on double routings to rail transit and a multi-objective programming model was developed to minimize the total cost, combining the passengers’ total travel time loss and the operator’s total operating cost. Based on the analysis of an urban rail transit train plan and users’ travel expenses, Deng et al. [
19] developed a multi-objective optimization model to reduce passengers’ general travel expenses and increase the operator’s benefits, simultaneously. Codina et al. [
20] presented a model balancing construction costs, operational costs and passengers’ benefits. The mode optimized users’ travel times and designed a public transit network system using the traditional approach of transport demand coverage in bimodal scenarios of operation. An and Zhang [
21] constructed a mixed integer programming model to minimize the cost while improving transit service. The model composited bus-holding and stop-skipping strategies to obtain a real-time optimal strategy and was solved by a Lagrangian relaxation algorithm. The study found a better balance between the quality of the transit service and the operation cost.
In some special cases, a mixed operation of double routing can also be used. Considering rotating maintenance, rolling stock circulation plans and a long-term maintenance policy, Canca et al. [
22] proposed a mixed integer programming model to minimize wasted capacity caused by train empty movements and to balance the workload of the maintenance operation. Cadarso et al. [
23] presented an integrated model for the recovery of the timetable and the rolling stock schedules, so that the recovery schedules could be easily implemented in practice and the operations could quickly return to the originally planned schedules after the recovery period. Cadarso et al. [
24] proposed an approach synthetically considering the influence of passengers’ behavior, the timetable and rolling stock. It combined two models: one which was an integrated optimization model for the timetable and rolling stock and another which was a model for the passengers’ behavior. The proposed approach could quickly find the best solution in two steps and with a very good balance between various managerial goals.
In summary, a large number of double routing operation models have been developed and the main optimization is to determine the return site and the departure frequency. On the basis of the aforementioned models, the purpose of this study is to develop a double-routing optimization model to minimize the objective function combining the passenger waiting time and the wasted capacity on the balanced scheduling mode, so as to obtain better capacity matching.
Compared with previous studies, this study presents the following differences:
Compared with Deng et al. [
19], in which the turn-back station cannot be changed, the model established in this manuscript is used to first make decisions on the turn-back station.
Compared with Deng et al. [
19] and Chang et al. [
18], this manuscript determines the routing which the train should be operated on at
xi time on the basis of considering the departure frequency and train capacity. Fleet size is not considered in this manuscript, because it changes infrequently in practice.
Compared with Deng et al. [
19] and Chang et al. [
18], the single-objective model is obtained by weighted summation with two different dimensions of passenger waiting time and wasted capacity in the original model and the nonlinear integer programming model is solved based on genetic algorithms.
Compared with Deng et al. [
19] and Chang et al. [
18], the relationship between the total passenger flow and the train’s optimum headway is obtained in this study, which makes it unnecessary to calculate the section passenger flow and find out the maximum section passenger flow specifically. In additional, this manuscript simplifies the algorithm for calculating departure frequency and analyzes the conditions for the double-routing operation.
Table 1 shows the detailed feature comparison between closely related studies.
By optimizing the train operation plan to meet the demands of different sections’ passenger flow, the cost of transportation enterprises is reduced, under the premise that the service quality is little affected, and the efficiency of the trains used is improved. At the same time, this operation mode can alleviate wasted capacity and make the operation organization economically reasonable.
The rest of this paper is organized as follows. In
Section 3, we provide the problem statement. In
Section 4, we formulate an integer optimization model with the objectives of passenger waiting time and wasted capacity. In
Section 5, we design a genetic algorithm (GA) to solve the formulated optimization model. Based on real-world passenger and operation data from Beijing Metro Line 4, in
Section 6 we conduct a numerical example.
Section 7 concludes the work.
3. Problem Statement
Transit timetable construction at the route level is generally based on the maximum observed passenger flow. In current practice, this maximum passenger flow is observed in a specified time period. It is divided by a load standard (desired occupancy, e.g., number of seats) to obtain the number of departures required [
25]. At present, urban rail transit is mostly operated on a single long routing mode, as shown in
Figure 1a and the train graph in
Figure 1b. With the city gradually extending outwards, there are metro lines connecting suburbs at both ends of the urban area. The passenger flow in the urban area is very large but in the suburbs the flow is relatively small at both ends of the line, so it is suitable to use the double-routing operation mode, as shown in
Figure 1c and the train graph in
Figure 1d. In addition, there are metro lines connecting the urban central area to the suburbs and similarly, the passenger flow in the urban area is much larger than the passenger flow in the suburbs, so it is suitable to use a simpler double-routing operation mode, as shown in
Figure 1e and the train graph in
Figure 1f.
A brief description of the physical urban rail line and rolling stock is shown in
Figure 2. The solid arrow stands for the long routing and the dotted arrow represents the short routing. The solid lines are the simple expression of the physical urban rail tracks. The reversing tracks have been built at station
a, consequently, the train operated on the short routing can change the travel direction.
The train routing plan is an important part of the train operation plan in urban rail transit. Its formulation should be based on the full analysis of passenger flow, considering the operation organization conditions. It will directly affect the formulation of the train timetable and the rolling stock operation plan. The motivation for the double-routing operation is to promote a better match between capacity and demand. Passenger demand may vary at different times or sites.
Figure 3 illustrates the spatial distribution of passenger flow and the capacity usage on each section. The spatial distribution of passenger flow reflects the difference in passenger flow in different directions and locations of the line. The usage of capacity in a single long routing is shown in
Figure 3a and we can find that there is a lot of waste of capacity in many sections. In support of the idea that operators should use the double-routing operation and reduce the number of trains in sections with less passenger flow, the capacity usage on the double routing is shown in
Figure 3b, in which the usage of capacity is improved on some sections.
Since many passenger demands can be met by a trip following either a full-length or short-turn pattern, schedule coordination between the patterns is essential [
3]. The critical issue in short-turn service is to determine the turn-back point and the route schedule to balance passenger loads among the trips and to minimize the total fleet size and passenger waiting times. The train plan of urban rail transit under the multi-routing mode can be divided into three parts: train formation, train operation periods and the corresponding train counts of each routing in each period [
19]. The design of strategies that guarantee reasonable user waiting times with small increases in operation costs is now an important research topic [
10,
26].
Through the above analysis, we construct a double-routing optimization model to determine the turn-back station of the short routing according to the uneven distribution of passenger flow. Then, we decide to operate on either the long or short routing at a particular time, based on the actual passenger flow. The trains on the long routing will turn back at the terminal, while the trains on the short routing will turn back at the middle station obtained by the model decision.
5. Solution Procedure
The determination of efficient routes and schedules in public transport systems is complex due to the vast search space and the multiple constraints involved [
27,
28,
29,
30,
31]. It can be seen from the above statement that the established model consists of the passenger waiting time and the wasted capacity, where minimizing the passenger waiting time is a nonlinear integer programming problem and the minimization of wasted capacity is a general linear mixed integer programming problem. To solve the general mixed integer model, the common solutions include the branch and bound method and the cutting-plane method; at the same time, there are also some stable solvers, such as WebSphere ILOG CPLEX and mixed integer linear programming solver-IPSOLVE. Considering the actual situation of urban rail transit (e.g., multiple stations and large passenger flow), we select the CPLEX solver to try to solve the problem at first but unfortunately it takes a long time [
32]. It can take up to half an hour for an optimal solution to optimize 35 stations and will be detrimental to the further development and expansion of the follow-up study. In addition, the solver does not solve the nonlinear optimization. Since with the developed double-routing optimization model it is difficult to find a good solution using the method of exhaustion or traditional optimization algorithms, we design a genetic algorithm (GA) in this study to solve the developed model, which greatly improves the solving efficiency and is more conducive to the expansion of the follow-up research. GA is a random global search algorithm and optimization method for seeking high-quality solutions. Due to its advantages of strong robustness, extensive generality and high efficiency, GA has been widely adopted for solving many urban rail optimization problems [
16,
33]. The basic steps of GA are coding, initial population generation, fitness evaluation, selection, crossover and mutation, as shown in
Figure 4.
Step 1. Chromosome coding: The double-routing optimization model has
f + 1 variables. The first
f decision variables are the train operation plan,
x1,
x2,
…,
xf; when
xi takes the value of 1, it means the train will operate on the long routing at time
xi; when the value is 2, it means the train will operate on the short routing at time
xi. The
f + 1 decision variable is the index of the return station in the short routing. Therefore, the chromosome can be divided into
f + 1 parts for coding, that is, one chromosome has
f + 1 genes. We convert decimal variables into binary genes. In this way, a solution to the model is encoded as a solution of the GA. As shown in
Figure 5, this chromosome represents that at time
x1 the train should operate on the long routing, at time
x2 the train should operate on the short routing and at time
xf the train should operate on the short routing. The index of the return station in the short routing is No.16.
Step 2. Initialization: According to the constraints of departure frequency and the return station in 18, initialize the population p0, generate n0 initial solutions randomly and ensure that each initial solution generated is a feasible solution.
Step 3. Evaluation function: Fitness evaluation is used to measure the quality of individual candidates. The GA is usually used to find the maximum value of the model; however, in this study we need to find the minimum value of the objective function of the double-routing optimization model. Here, we use the function f(x) = M − z(x) as the evaluation function, where M is a constant that ensures that the value of f(x) is positive. If the value of f(x) is bigger, the corresponding solution is better.
Step 4. Selection process: In this paper, the best reserved selection method is adopted. First, roulette is used for selection and then the individuals with the highest fitness in the current population are completely copied to the next generation. The cumulative probability that individual
xi is inherited to the next generation is
Step 5. Crossover process: The crossover is the most important step in the GA, which could generate new chromosomes in the population. This paper uses two points of crossover. First, two chromosomes are paired randomly and serve as parent chromosomes, which are selected for the new population according to the crossover probability. If the new chromosome meets all constrains, take it to replace the parent chromosome. Otherwise, keep the parent chromosome. The crossover process is shown in
Figure 6.
Step 6. Mutation process: The crossover process results in local optimum solutions. Thus, a mutation process is used to escape local optimums. For each chromosome in the population, we first determine whether it can be selected for the mutation operation by mutation rate. For the selected parent chromosome, randomly select a code of the chromosome. If the selected code is 0 (or 1), take 1 (or 0) to replace it. If the new chromosome meets all constrains, take it to replace the parent chromosome. Otherwise, keep the parent chromosome. The mutation process is shown in
Figure 7.
6. Numerical Example
In this section, the GA is coded in Matlab programming language to verify the validity of the double-routing optimization model. This work mainly studies one direction. Beijing Metro Line 4 is a trunk line connecting the north and south areas of the western part of Beijing. We take Beijing metro line 4 as the research object, which is 50 km long and has 35 stations. First, we mark Tian’GongYuan Station as No. 1 and the Biomedical Base Station as No. 2, continuing the marking up to No. 35, in turn. The route diagram of Beijing Metro line 4 is shown in
Figure 8 and the daily travel demand on line 4 is shown in
Figure 9.
Given the OD data for Beijing metro line 4 for the whole day of 9 July 2018, this paper selects the OD data from 8:00 to 9:00 to form the OD matrix, as shown in
Figure 10. The uneven distribution of OD is visually reflected by different shades of color.
The passenger flow in each section is shown in
Figure 11. During the research period, there were a total of 88,645 trips from Tian’GongYuan Station to AnHeQiao North Station and 49,681 trips from AnHeQiao North Station to Tian’GongYuan Station. The uneven passenger flow distribution in both directions is caused by the large number of morning commuters from the suburbs to the city. We choose the direction with high passenger flow from Tian’GongYuan Station to AnHeQiao North Station as the optimization target. It can be seen from
Figure 10 that the maximum passenger flow is from CaiShiKou Station to XuanWuMen Station—in this section there were 35,391 passengers bound for AnHeQiao North Station and 10,273 passengers bound for Tian’GongYuan Station. For this reason, we choose [
a,
n] as the short routing.
According to the code for the design of the metro and combining the actual operation characteristics of the metro systems, the equipment operating data and some parameters in the model are shown in
Table 2.
This paper implements GA with Matlab programming to optimize the double-routing optimization model. The relevant parameters of the GA are shown in
Table 3.
When the objective is to minimize passenger waiting time, the decision variables are obtained as shown in
Table 4, which indicates that all trains should run on the long routing; at the same time this will cause a certain wasted capacity. When the objective is to minimize the wasted capacity, the decision variables are received as shown in
Table 5, indicating that there are 15 trains running on the short routing and 9 trains on the long routing in an hour; this will cause the minimum wasted capacity and a greater passenger waiting time. A specific comparison of the effects on different objective functions is shown in
Table 6.
It can be seen that taking passenger waiting time or wasted capacity as the objective function is not enough, since this will either cause a certain wasted capacity or seriously affect the service quality. Therefore, the objective function combining passenger waiting time and wasted capacity is very necessary.
We substitute the parameters into the double-routing model and find the solution. The variation tendencies of average fitness, optimal fitness, the average distance between individuals and scores in the iteration process of the genetic algorithm are shown in
Figure 11 and thus we obtain the current best individual.
It can be observed from
Figure 12 that in the iteration process of the genetic algorithm, the average fitness and the optimal fitness show a downward trend and so do the scores and average distance between individuals. The variation tendencies of average fitness and optimal fitness show that the algorithm has good convergence. When the objective function takes the minimum value, the values of the decision variables are as shown in
Table 7. The index of the return station in the short routing is No. 10, namely XiHongMen Station. The short routing is [XiHongMen Station, AnHeQiao North Station].
The train timetables before and after optimization are shown in
Figure 13;
Figure 14. The x-coordinate indicates the departure times of trains and the y-coordinate indicates the stations along Metro Line 4. It can be seen from
Figure 13 and
Figure 14 that for trains operated on balanced scheduling mode, every 2.5 min a train will be sent from AnHeQiao North Station and Tian’Gong Yuan Station. The blue line parallel to the X-axis in
Figure 14 represents XiHongMen Station, which is the turn-back station of the short routing.
In
Figure 14, there are four trains that turn back at XiHongMen Station; the rest of trains are operated on the same timetable as before the optimization. We compared the effects before and after optimization obtained by using the double-routing optimization model and the specific contents are shown in
Table 8.
It can be observed from
Figure 13,
Figure 14 and
Table 8 that the wasted capacity of Beijing Metro Line 4 decreased by 9.5% with the optimized train operation plan, while the passenger waiting time increased by 4.5%, which can verify that this optimization model is very effective. We take the passenger flow in two directions into consideration and substitute them into the model to obtain the result, as shown in
Table 9. In the best found train timetable, as shown in
Figure 15, there is one train that turns back at XiHongMen Station, while the rest of trains are operated the same as before the optimization.
As can be seen from
Table 9, when two directions are taken into consideration, the index of the return station does not change but the departure frequency on the short routing decreases and on the long routing it increases accordingly.
The departure frequency on a single long routing can be calculated by the maximum section passenger flow, which is our consensus. Through the in-depth study of this model, we obtain the relationship between the total passenger flow and trains’ optimum headway, which makes it unnecessary to calculate the section passenger flow and find out the maximum section passenger flow specifically. To some extent, this facilitates our calculation. As shown in
Figure 16, the horizontal axis is the ratio of the total passenger flow to the maximum passenger flow; the maximum passenger flow here is calculated according to the maximum train operated on the single long routing and the load factor is 130%. The vertical axis is the optimal headway.
The following part will simulate some different compromise solutions and study some influencing factors of the model, such as the location of the return station in the short routing, the number of trains on the short routing, the ratio of passenger flow in the short routing and the upper limit of the load factor, in the hope of reaching recommendations for metro trains operated on the double routings.
The location of the short routing’s return station will affect the ratio of passenger flow and the short routing’s turnover time, which will have an impact on the optimal train routing plan and the passengers’ total waiting time. We set other variables as fixed and change the index of the return station. The optimal train routing plan and the influence of the return station’s location are shown in
Table 10.
From
Table 10, it can be seen that the departure frequency on long and short routings changed when the location of the turn-back station moved. When the short routing is [10,35], the departure frequency on the short routing is 4 tr/h; at this point, we select
h = 35 as the origin station and fix it, moving the terminal outwards or inwards and the frequency on the short routing reduces. When the short routing is [
10,
33], the departure frequency on the short routing is 5 tr/h; we select
h = 10 as the terminal and fix it; when the origin station is moved outwards or inwards, the frequency on the short routing reduces. It can be found that when the short routing is [
10,
30] and the short routing’s departure frequency is 4 tr/h, the combined cost is at the minimum but at this time, there are two stations that need to be reconstructed. If we intend on achieving the optimization purpose by reconstructing only one station, the optimal routing plan is [10,35] and the departure frequency on the short routing is 4 tr/h.
When the short routing is fixed, the proportion of trains on the short routing will directly affect the passenger’s total waiting time.
Table 11 and
Figure 16 show the impact of the different train proportions on passenger waiting time. At this time, the total number of trains is 24 and the short routing is [10,35].
The service level is closely related to passenger waiting time. From
Table 11 and
Figure 17, it can be seen that the total passenger waiting time increases, with the number of trains on the short routing increased. In order to ensure the service quality, the number of trains on the short routing cannot be too many. Therefore, we can determine the maximum number of trains on the short routing based on the service quality requirements of the metro operation company.
The passenger flow in the short routing refers to the flow of passengers whose origin and destination are both within the short routing, that is, In this case study, O ∈ [10,35] and D ∈ [10,35]. In order to study the influence of the ratio of the short routing’s passenger flow, the original data are processed by increasing or decreasing the passenger flow in [10,35] by 10% ,20% or 30%, successively. The different passenger flow distribution patterns are shown in
Figure 18, which are +30%, +20%, +10%, −10%, −20% and −30% different from the original passenger flow, respectively. The optimal train routing plans under different passenger flow conditions are shown in
Table 12.
By changing the passenger volume, it can be found that when the passenger flow in [10,35] is increased, the departure frequency is first determined according to the maximum section passenger flow and then the short routing with higher passenger flow is re-selected to complete optimization. When the passenger flow increases by 10%, the departure frequency increases to 27 tr/h. The departure frequency cannot keep increasing, due to the constraints of safe operation and the signal system, therefore, when the passenger flow increases by 20% or even 30%, it is determined as 30 tr/h. When the passenger flows in [10,35] are decreased by 10% and 20%, the volumes in the short routing are only 5.4% and 3.2% higher than the average section passenger volume. The optimal train routing plans obtained by the model are the same as the single long routing, indicating that in these cases, metro trains are not suitable for operating on the double routing. When the passenger flow in [10,35] decreases by 30%, the passenger flow of each section is basically the same. The wasted capacity caused by the single long routing will be relatively low and the requirement can be satisfied by trains operated on the single long routing. Therefore, we infer that when there are some sections whose passenger flow is more than 6% higher than the average section passenger flow, the double routing model can be considered.
The change in the upper limit of the load factor will directly affect the value range of feasible solutions. In
Table 13, the optimal train routing plans under different load factors are shown.
The upper limit of the load factor will affect the passenger waiting time and the train departure frequency. When the load factor declines, the capacity of each train will decrease and passenger flow can be satisfied by increasing the departure frequency. With the increase of the load factor, the departure frequency on the short routing increases, while the frequency on the long routing decreases correspondingly. The reduction of the departure frequency on the long routing will reduce the waste of transportation capacity but at the same time will lead to the increase of passenger waiting time on non-double-routed sections, which is a mutual restriction process. When the upper limit of the load factor is 120%, the obtained short routing is very close to the complete long routing. The above experimental results basically agree with the theoretical analysis and the actual situation, which validates the feasibility of the optimization model.