A Two-Stage Queue Model to Optimize Layout of Urban Drainage System considering Extreme Rainstorms

Extreme rainstorm is a main factor to cause urban floods when urban drainage system cannot discharge stormwater successfully. This paper investigates distribution feature of rainstorms and draining process of urban drainage systems and uses a two-stage single-counter queue methodM/M/1 → M/D/1 to model urban drainage system.Themodel emphasizes randomness of extreme rainstorms, fuzziness of draining process, and construction and operation cost of drainage system. Its two objectives are total cost of construction and operation and overall sojourn time of stormwater. An improved genetic algorithm is redesigned to solve this complex nondeterministic problem, which incorporates with stochastic and fuzzy characteristics in whole drainage process. A numerical example in Shanghai illustrates how to implement the model, and comparisons with alternative algorithms show its performance in computational flexibility and efficiency.Discussions on sensitivity of fourmain parameters, that is, quantity of pump stations, drainage pipe diameter, rainstorm precipitation intensity, and confidence levels, are also presented to provide guidance for designing urban drainage system.


Introduction
A storm drainage system is a critical infrastructure for a city to support its development, but many cities in China are suffering severe flood problems because of inadequate investment in their drainage system [1,2].An urban flood may seriously ruin buildings, block traffic, and even threaten resident's life.Although there are many factors to cause urban floods, inadequate drainage capacity and improper drainage layout are main factors.Human behavior as well as economy varies in different districts, whose hydrological characteristics change, and a rainstorm would cause different consequences [3].
An urban drainage system usually consists of groundwater convergence points, drainage pipes, and stormwater pump stations and outlets.Its drainage capacity is determined by complicated combination of the above components.The draining operation process can be regarded as a serial two-stage queuing system with three main activities: in the first queue stage, the drainage system is designed to collect stormwater runoff from the road surface and right-ofway.Stormwater collection is here accommodated through the use of drainage inlets, which serve as the mechanism whereby surface water enters underground storm drainage pipes.Drainage inlet locations are often established by the roadway and on side streets to reduce the spread of water onto the roadway surface.Upon reaching the drainage pipes, stormwater is conveyed along to its discharge point.In the second queue state, the drainage system is to discharge stormwater from the drainage pipes by stormwater pump station to an adequate receiving body such as a stream, creek, river, lake, or other piped systems.A stormwater pump station is often required to drain depressed sections of drainage pipes where gravity drainage is impossible or not economically justifiable.
Many researchers have studied the queuing theory of serial multiple service stations.Paper [4] presented a twostage queue //1 model in solving the state probabilities of the multiqueue problem.Paper [5] developed an  [] /( 1 ,  2 )/1 retrial queue model to solve a repeated drainage problem.Paper [6] focused on a two-stage queuing network for outfitting process in shipbuilding.Paper [7] extended the cascaded queuing system to infinite servers queue system to solve logistics industry problems.It is a reasonable approach to adopt serial multiple queue model for this urban drainage system layout problem.Many researchers also developed queue models to optimize and design facility layout.Paper [8] presented an innovative approach to solve a layout problem with a multiobjective nonlinear mixed integer programming model considering // queening congestion.Paper [9] developed a queuing model of ambulance service to calculate key performance indicators which both managers and patients focused on.Paper [10] proposed a multipriority coverage layout model based on queuing theory.
But a facility layout system often faces uncertain factors which might result in stochastic demands; how to use queue models to design a facility layout in an uncertain environment attracts many researchers.Papers [11], [12], and [13] considered influences of stochastic factors on the multiobjective queuing location-allocation problem, respectively.Papers [14][15][16] discussed the uncertain problems of fuzzy factors on location by fuzzy queuing models.
When designing a stochastic facility system, its objectives and constraints are stochastic; how to solve this kind of nonlinear programing problem is a challenge which few literatures observed.Paper [17] developed a queuing model considering stochastic custom demands and unmovable service counter in a facility layout problem and then solved it by genetic algorithm with expectation value.Paper [18] established a biobjective model for bidirectional facility network under uncertain supplies and then presented a solution method combined with queuing theory and fuzzy programming.
However, previous studies mainly focus on such problems as optimization of queuing layout routines or assignment of customs; they have never tried to solve complex problems such as the following: (1) effects of extreme demands from customers, for example, extreme climate; (2) facility layout as a strategic decision problem because a drainage facility often operates for a long time; (3) priority of custom demands in various areas where a rainstorm might have different damage and losses in different urban neighborhoods.
This paper considers characteristics of extreme rainstorm and urban flood, stochastic demand for drainage, drainage system capability, fuzzy sojourn time of the system, and draining priority in different areas and then discusses the problem of drainage facility layout under extreme rainstorms.A biobjective, drainage system cost and drainage sojourn time model will be developed, and a stochastic and fuzzy generic algorithm is presented to optimize stormwater impacts on drainage facilities.
The remainder is structured as follows.The mathematical modeling of urban drainage system layout under extreme rainstorm is introduced in Section 2, including the identification of parameters and decision variables, assumptions, methodology for stochastic and fuzzy problems, and constraints.A solution method of improved genetic algorithm for nondeterministic problems with hybrid constraints of randomness and fuzziness is described in Section 3.An example illustrating the parameter setting and flexibility of the proposed algorithm is provided in Section 4. Sensitivities of several parameters and their patterns are discussed in Section 5. Finally, some conclusions are provided in Section 6.

Mathematical Modeling
An urban drainage system is designed to collect stormwater runoff from the roadway surface through inlets, to convey it through drainage pipes, and to discharge it to a receiving body by pump stations.The draining process is regarded as a stochastic two-stage queuing system in this paper, where stormwater, as customers, arrives at the drainage system in uncertain time and stochastic quantity.The stochastic and fuzzy features increase complexity in designing urban drainage system.A mathematical model of urban drainage system with stochastic and fuzzy parameters is developed as follows.

Parameters and Decision
Variables.The model parameters and their notations and decision variables are listed in Notations.
Both  and  are set of discrete points in a draining district. is determined by   ,   ,   , and ℎ  .A smaller  means the construction and operation cost of urban drainage system is less.  is drainage pipe length which is an equivalent distance of drainage pipe from point  to pump station .  is water capacity of drainage pipe whose diameter is . is total sojourn time of stormwater collected, conveyed, and discharged by urban drainage system, determined by the longest sojourn time max{ t }.The smaller t is, the more efficient the drainage system is.Sojourn time t is a fuzzy value, and its mean value is affected by   ,   , and   .  is demand rate at which stormwater at point  receives service from drainage pipe, and it is relevant to precipitation intensity   at point .
There are three decision variables,   ,   , and   , in this model.  and   are 0-1 variables, and   is a continuous variable.  indicates whether pump station  is constructed at point  or not,   indicates whether pump station  provides service to point  or not, and   is service ability, namely, draining capability, of pump station .

2.2.
Assumptions.An extreme rainstorm often shows such characteristics as outburst, intensive precipitation, uneven geographical distribution, and short duration.If such intensive stormwater is not drained out in time, flood or waterlog will appear.When designing a drainage system, all rainfall, equivalent to precipitation in its covered area, is a critical factor.Although precipitation in urban district can be obtained from historical weather data, it is hardly predicted accurately because of its uncertain and random weather conditions.The precipitation duration and intensity are also regarded as stochastic variables.Here are main assumptions in modeling urban drainage system in this paper.
Assumption 1. Precipitation intensity of an extreme rainstorm in an urban district is subject to a stationary Poisson distribution and the duration is stochastic [19,20].
When precipitation intensity   at point  is subject to a Poisson distribution, its distribution function  of precipitation intensity   is The parameter  in the distribution function  could be determined by the maximum likelihood estimation.If there is a population  and sample  1 ,  2 , . . .,   , then its maximum likelihood function [21] is The parameter  would be replaced by the maximum likelihood estimation τ: This distribution function adaptability then can be tested by chi-square distribution method which has been broadly applied in fit test [22,23].The test hypotheses are as follows: All values Ω from  are divided into  subsets:  1 ,  2 , . . .,   , which are pairwise disjoint.Observation of samples  1 ,  2 , . . .,   reveals that   ( = 1, 2, . . ., ) belongs to   and its occurrence frequency is   /, where  −1 and   are, respectively, the lower and upper bounds of group   .Then, the test probability p and the fitting value  2 are, respectively, obtained: There are  parameters to be estimated, and their confidence level is .If  2 ≤  2  ( −  − 1), then the hypothesis  0 will be accepted and its expectation and variance are   =  and  2  = , respectively; otherwise, the hypothesis  1 will be accepted.Assumption 2. All stormwater which is drained by drainage pipes comes from inlets, and its volume is equal to quantity of all precipitation.
Stormwater quantity at point  is multiple of its precipitation intensity, precipitation area, and precipitation duration; that is, where   is the rainfall area covered by drainage pipes (pipe diameter is   ) and   is precipitation duration.
Assumption 3. Stormwater flows in separate queues inside drainage pipes and its flowing sequence will not change until arriving at its pump station.The arrival time of stormwater which flows from drainage pipes to its pump station is stochastic, and only the nearest pump station will provide service to it.
Assumption 4. Urban drainage system consists of several drainage pipes and pump stations, and the whole draining procedure in which stormwater is processed by drainage system is regarded as a serial two-stage queue system, namely, //1 → //1 queue system, subject to first-come firstserved (FCFS) order.
The service counter in the first stage of this queue system is the drainage pipe, and its serviced object is stormwater runoff collected by inlets and its service capacity is drainage pipe's water capacity.In the second stage, the service counter is the pump station and its serviced object is stormwater conveyed inside drainage pipes which finally is discharged into rivers or other bigger pipes, and its service capacity is pump station's operation capacity.
The service time of drainage pipe is regarded as an independent random variable which is subject to a negative exponential distribution [24].The service capacity of drainage pipe  is   , which is relevant to its diameter: where  is the pipe diameter and V  is the average flowing velocity inside drainage pipe .
A drainage pipe  only receives service from one pump station .If service capacity of the pump station  is   , the service time and the arrival time in two stages are independent, the demand rate of stormwater at point  is   , and the average arrival rate of stormwater arriving at pump station  is   , then, service intensities in two stages are  1 and  2 , respectively: Assumption 5.Each pump station in the urban drainage system is independent but each one has an independent drainage capacity.The sojourn time of a pump station discharging stormwater is fuzzy.
A deterministic sojourn time   of stormwater collected from point , conveyed by drainage pipe and discharged by pump station  in urban drainage system, can be derived: However, there are many factors affecting the sojourn time; it is regarded as a fuzzy variable in this paper.Since triangular fuzzy number has been proven to be excellent by previous literatures in solving fuzzy simulation problems [25,26], the sojourn time is regarded as a triangular fuzzy number and it can be transferred to a possibilistic value by this method.The total sojourn time of stormwater from point  processed by pump station  can be represented by a triangular fuzzy number t : t = (  ,   ,   ) .
Its fuzzy membership function is Considering confidence level of fuzzy number, its -cut set is The upper possibilistic mean value   ( t ) and the lower possibilistic mean value   ( t ) of drainage sojourn time t in -cut set are obtained: The clear possibilistic mean value of the fuzzy numbers is obtained from   ( t ) and   ( t ) [27].Therefore the possibilistic mean value ( t ) of drainage sojourn time t is as follows: The possibilistic variance ( t ) of the total drainage sojourn time is as follows: Assumption 6.If a region at point  is more important than other regions, a smaller restriction value   will be applied to reduce its sojourn time of stormwater.  is a stochastic variable and subject to normal distribution.When there are  experts, each one ranks a restricted sojourn time   of stormwater at point ; then, its expectation   and variance  2  can be obtained, and its restriction value   of sojourn at point  can be described by this normal distribution function (  ,   ).
One of the service time constraints is the regional drainage priority based on their importance.The same rainstorm has varied impacts on different districts because their characteristics such as population, traffic, and economics are totally different.In order to reduce potential flood or waterlog in key regions such as commercial centers, major communities, busy roads, and political and culture centers, a regional priority is introduced to adjust their drainage sojourn time as well as drainage service capability in different regions.An expert ranking method or weighting method is often used in choosing key regions.So several experts can be asked to choose a constraint value   to adjust regional sojourn time.
Assumption 7. Total cost of urban drainage system includes (1) construction cost of drainage pipes, (2) construction cost of pump stations, (3) equipment cost of pump stations, and (4) operation cost of drainage system.Among them, construction and equipment cost of pump stations is supposed to be proportional to its drainage capacity, and pipe's construction cost is proportional to its length.

Methodology for Modeling a Stochastic and Fuzzy
System.An urban drainage system incorporates randomness and fuzziness because of its uncertain environment, so its model consists of mixed conditions with a lot of random and fuzzy parameters.After considering this kind of mixed conditions, a methodology to model a stochastic and fuzzy system is presented as follows: where Pos{⋅} represents possibility of an event in { } and Pr{⋅} represents probability of an event in { }.  is a decision vector,  is a stochastic parameter vector, and η is a fuzzy parameter vector.(, , η) is an objective function, and   (, , η) is a constraint function.

A Biobjective Model of Urban Drainage
System.The two objectives in this biobjective model of urban drainage system are total sojourn time  of stormwater and total cost .The total sojourn time  is the maximum of sojourn time   of stormwater from all points  ( ∈ ).The total cost  consists of construction cost, equipment cost, and operation cost.
Considering stochastic rainstorm demand, fuzzy sojourn time, and its restriction value, the biobjective model with random and fuzzy constraints for urban drainage system is developed: Pr ∑ The objective function ( 16) is to minimize total sojourn time of stormwater processed by drainage system, and constraint function (18) represents the possibility of total sojourn time including waiting time and service time which is not less than  1 .The objective function (17) is to minimize the expected value of total cost that the stormwater is processed by drainage system, and constraint (19) is to ensure cost possibility is not less than  2 and denotes the total cost components.
Constraint (20) is to ensure probability of total sojourn time of drainage system subject to its restriction value is not less than  1 and its possibility is guaranteed to be not less than  2 .Constraint (21) denotes a method to estimate length of drainage pipes.Constraint ( 22) is to ensure stormwater will not exceed the maximum water capacity of drainage pipes.Constraint (23) ensures that probability of balance between service ability of pump stations and drainage demands is not less than  3 .Constraint (24) denotes that stormwater demanding service of drainage system appears only after establishment of a pump station at location  is finished.Constraints (25) and (26) indicate that stormwater at point  is only served by its nearest drainage pump station.Constraint ( 27) is a given quantity of pump stations to be constructed in this system.Constraint (28) denotes the average arrival rate of stormwater to pump station  in unit time.Constraint (29) is property of two decision variables.

An Improved Genetic Algorithm for Stochastic and Fuzzy System
Since a stochastic and fuzzy system demonstrates complex nonlinear characteristics in its objectives and constraints which are not easy to be converted into deterministic variables, it is difficult to solve a large-scale nonlinear problem by traditional algorithms.Fortunately, many heuristic algorithms such as genetic algorithm have been developed by researchers to solve deterministic problems [28] as well as nonlinear problems [29].
In order to apply genetic algorithm in stochastic and fuzzy urban drainage system, its nondeterministic variables would be sampled based on probability and possibility distribution method and then a simulation model could be developed [30].After improvement of genetic algorithm, nonlinear problems in urban drainage system would be solved.

Method of Simulating Nondeterministic Variables. For a given decision vector
In order to verify the above equation, a fuzzy vector η generates evenly an  0 which satisfies ( 0 ) ≥  2 ; namely,  0 is extracted from level -cut set of η, or  0 is extracted from a super geometry body Ω which includes level -cut set.The condition ( 0 ) ≥  2 determines whether  0 is accepted or not.
It is supposed that the integer part of  1  is   ,   = int( 1 ), and the   th biggest element of { 1 ,  2 , . . .,   } is  0 .If  0 ≤ , then  =  0 .After repeating  times of choices,  is regarded as the objective value at point .

Improved Genetic Algorithm for Urban Drainage System.
Genetic algorithm is a high parallel, random, and adaptive intelligent optimization algorithm [31].An improved genetic algorithm can simulate stochastic and fuzzy characteristics in its process of producing population, crossover, and mutation operations and calculating value of chromosome [32].An improved genetic algorithm is redesigned to solve urban drainage system as follows.
(1) Coding.The decision variables   and   are coded by a mixed coding method which is combined with floating point and binary code.The code length is  +  × , where ,  represent amount of stormwater inlets points and the amount of pump stations, respectively.
The (+1)th to the (+×)th digits in the code are binary numbers, which represent decision variable   .If   = 1, it denotes that stormwater will be discharged by pump station .If   = 0, it denotes that stormwater will not be discharged by pump station .Each chromosome can be decoded, and its genotype and phenotype can be exchanged.
(2) Initialization.Several random points are generated from the feasible domain of decision vectors when chromosome   is initialized.Then, its feasibility would be tested by a random and fuzzy simulation method.If it is feasible, the chromosome will be selected.Otherwise, a new random point should be generated until a feasible solution is obtained.After repeating  times, there are  initial feasible chromosomes   obtained.Target values of all chromosomes would be obtained by the random and fuzzy simulation method.
(3) Setting Fitness Function.A probability value is used in each chromosome   which is determined based on order of an evaluation function eval(  ).Roulette wheel selection method is applied to make the possibility of each selected chromosome be proportional to the adaptability of other chromosomes in the same population, meaning that a more adaptable chromosome would be selected firstly and to produce possibly more offspring.
The evaluation function eval(  ) is eval where  ∈ (0, 1) and usually  = 0.05. is ordered from good to bad according to values of objective functions.A better chromosome is obviously assigned a smaller number.
(4) Selection.Cumulative probability is calculated based on evaluation function eval(  ) of each chromosome   .
A random number   is generated from the interval [0,  popsize ].If  −1 <   ≤   , then the th chromosome   is selected, and  times of duplicated chromosomes are available after repeating  times.
(5) Crossover.The probability of crossover operation is   and the following process is repeated from  = 1 to popsize: A random number   is generated from the interval [0, 1].If   <   , chromosome   is selected as a parent to participate in crossover operation.After selection, the parents are named    ,  = 1, 2, . .., and they are divided into pairwise groups (  1 ,   2 ), (  3 ,   4 ), . ... If the number of parents is odd, one of them is deleted randomly.A random number  is generated from (0, 1) and crossover is operated in the number of the first  digits of each parent.Therefore, two children are generated from the first  digits.The crossover operation is expressed as follows: The numbers from the ( + 1)th digit to ( +  × )th digits of each parent are crossed as follows: A random genetic position is regarded as the cross point, where the same genetic positions at two parents are exchanged.Then, two children  +1 ,  +1 are generated from the (+1)th digit to (+×)th digit, respectively, in probability   .
Combination of   and  +1 and   and  +1 will generate two new offspring  and .The feasibility of offspring is tested by a random and fuzzy simulation method.If its test is feasible, it will be replaced by its parents.Otherwise, a new crossover operation will repeat until two feasible offspring are obtained or the iteration reaches a limit cycle.
(6) Mutation.The probability of mutation operation   and following process is repeated from  = 1 to popsize: A random number   is generated from the interval [0, 1].If   <   , then the chromosome   is selected as a parent to participate in mutation operation.
Mutation is operated in the number of first  digits of each parent as follows: A genetic position is randomly selected as the mutation position, its genetic value is replaced by a random value generated from [0, +∞), and the number of its first  digits is regarded as its offspring   .
Mutation is operated at the number of the (+1)th digit to ( +  × )th digit of each parent as follows: A genetic position is also randomly selected as the mutation position, where 1 is replaced by 0 and 0 is replaced by 1.Then, its offspring  +1 is generated at the ( + 1)th digit to ( +  × )th digit.
Combination of   and  +1 is to generate a new offspring .The feasibility of offspring is tested by a random and fuzzy simulation method.If its test is feasible, it will be replaced by its parents.Otherwise, a new mutation operation will repeat until a feasible offspring is obtained or the iteration reaches a limit cycle.(7) The above process of selection, crossover, and mutation will repeat until they meet a termination criterion.
Finally, all optimal solutions of decision variables, sojourn time, and drainage cost are obtained, respectively.

Illustrative Example
Since a stochastic and fuzzy problem is so complicated that it hardly validates the proposed algorithm by any direct approaches, after reviewing the methods in [33,34], this stochastic and fuzzy problem would be simplified as a simulated deterministic problem.In order to verify this algorithm's validity to solve the large-scale nonlinear complex model, a numerical example is chosen and several tests are validated and discussed.

Description of Example and Setting of Parameters.
A square district in Shanghai is chosen as a numerical example to illustrate how to optimize drainage system layout by this algorithm.The district is 3200-meter-long and 3200-meterwide, but it is suffering frequent extreme rainstorms which result in more losses.In order to reduce urban floods, local authority wish to rezone drainage area and redesign its drainage system soon.It is divided into 64 drainage blocks whose length is 400 meters, shown in Figure 1.Drainage block's centers are supposed to be drainage points and all groundwater inside a block only flows to its drainage point where an inlet is installed.A drainage pipe is laid underground to connect with an inlet and a pump station.All groundwater is collected by inlets, conveyed by underground drainage pipes to pump stations.Any drainage point could be a candidate position of pump station.Stormwater from one drainage point only flows to one pump station through a drainage pipe; its flowing direction would never change.
There will be eight stormwater pump stations to be constructed in this district.The goals of this project are to drain all stormwater within a minimal sojourn time and to minimize construction and operation cost of this project at 95% confidence level when an extreme rainstorm hits this district.
Other parameters of simulation settings are defined as follows:   = 1 yuan/m 3 ,  = 1200 mm, and   = 1.131 m 3 /m.  is a linear distance between point  and point .
Precipitation intensity   at point  follows a Poisson distribution, which is estimated by 10-year 24-hour statistical precipitation data of extreme rainstorms.The probability distributions of precipitation intensity at all points ( 1 ∼ 64 ) are shown in Table 1.
Then, the distribution of precipitation intensity is tested by chi-square distribution method.Taking the point  = 1 as an example, where  1 obeys a Poisson distribution of its parameter  = 7.60, all possible values Ω from  1 are divided into seven groups, marked as  1 = [0, 2],  2 = (2,4],  3 = (4, 6],  4 = (6, 8],  5 = (8, 10],  6 = (10, 12], and  7 = (12, +∞].The chi-square distribution test process includes calculation of   , p , ĝ , and   2 /ĝ  , shown in Table 2. After test, it is obtained that  2 = 7.8313 and  2 freedom degree is  −  − 1 = 5.At the confidence level  = 0.05,    The first 64 digits in a chromosome gene are the decision variable   .The rest digits, namely, the 65th to (64+64×64)th  digits, are decision variable   .When chromosome   is initialized, a random point is generated from feasible domain of decision vector and its feasibility is verified by the random and fuzzy simulation method.The evaluation function eval(  ) = 0.05(1 − 0.05) −1 ,  = 1, 2, . . ., 200.Each chromosome   is then operated in selection, crossover, and mutation with a given probability and its offspring's feasibility is verified by the random and fuzzy simulation method.The algorithm process will terminate after 1000 iterations.Finally, the optimal solution and the objective function values are obtained.

Main Results.
The optimal locations of eight pump stations in this drainage system are  10 ,  13 ,  23 ,  27 ,  37 ,  39 ,  50 , and  54 .Its layout of inlets, drainage pipes, and pump stations is shown in Figure 2.Each pump station connects with six to nine inlet points by drainage pipes.All drainage inlets, pipes, and pump stations form an urban drainage system which provides optimal draining service for this district.This drainage system can meet the urban drainage requirements at 95% service level, which means that two objectives, to minimize the total sojourn time and to minimize the total cost, are guaranteed to be fulfilled at not less than 95% level.
The optimal draining capacity of each pump station is shown in Table 4.
The optimal values of two objective functions are  = 1738.6min and  = 4841.74 million yuan, respectively.Their confidence levels are The total sojourn time t of stormwater in this drainage system is stochastic, and their results from 64 points are described by three numbers (  ,   ,   ), shown in Table 5.
The stochastic and fuzzy characteristics cause this system to vary in searching for optimal results in this case study.After repeating execution of this model for times, it reveals that probability and possibility of all optimal results can be guaranteed at more than 95% level.Therefore, it is verified that this model and improved algorithm are flexible and effective to solve complex nonlinear problems of drainage system.

Comparing Results by Alternative Algorithms.
After repeating 10 times of solving the test by the redesigned genetic algorithm (RGA), their results of average optimal solutions including total time , total cost , and running times are listed in Table 6.Its results are compared with alternative algorithms such as simulated annealing genetic algorithm (SA-GA) [35] and Tabu Search (TS) [36].Their parameters are set as follows: Population size  = 200, crossover probability   = 0.75, mutation probability   = 0.04, maximum iteration number Maxgen = 200∼2000, and TS length  = √ , where  is quantity of candidate sets.The performance improvement of RGA is measured by comparison of searched total time  and total cost  with alternative algorithms, respectively, as follows: The comparison reveals that RGA is better than SA-GA in searching optimal solutions of total time  and total cost  with max performance improvement of 8.43% in  and 3.72% in  and average performance improvement of 4.32% in  and 2.55% in , and RGA is better than TS with max improvement of 5.77% in  and 2.68% in  and average performance improvement of 3.06% in  and 1.67% in .It is validated that RGA has the best search ability to find out optimal solutions.
While comparing the running times of algorithms, RGA needs less running time than TS but needs slightly more running time than SA-GA.For example, RGA, SA-GA, and TS needs 173 s, 162 s, and 187 s, respectively, in solving Case 1, which reveals that RGA needs only 11 s more than SA-GA but 14 s less than TS.Because RGA is s single step algorithm redesigned for urban drainage layout system but TS is a twostep algorithm, RGA is generally faster than TS in searching optimal solutions.Although RGA needs more running time, 11 s, this running time is acceptable.

4.4.
Validating the Proposed Algorithm.In order to validate the proposed algorithm, RGA is also reasonable and effective in solving fuzzy problems besides their stochastic features; Case 1 is chosen as a numerical test which is modified according to fuzzy features of urban drainage system.It is supposed that the poison distribution parameter  of precipitation density is 7.60 mm/h.After review of [37], sojourn time   is a triangle fuzzy number by simply extending it at scale  as follows: t = ((1 − )  ,   , (1 + )  ).
Parameters of RGA in this test are the same as those in the above cases, and its repeat number is 100.Running time of RGA is nearly 17685 s, 4.9 hours, which is about 100 times of deterministic Case 1.The results after ten operations are shown in Table 7.Since precipitation is stochastic and sojourn time is fuzzy, neither total sojourn time nor total cost of urban drainage system is deterministic.
This demonstrates the following results: (1) Layout of pump stations is exactly the same in all optimal solutions of 10 tests; (2) maximum total cost in this test is 2500.37 million yuan, minimum total cost is 2492.71 million yuan, and mean total cost is 2496.08 million yuan, and cost deviation is 0.31%; (3) maximum total draining time in this test is 3887.5 minutes, minimum total draining time is 3837.4minutes, and mean The iteration convergence graphs of total cost and total time at the 5th test, best close to the mean result, are observed and shown in Figures 3(a) and 3(b).After nearly 15 iterations, RGA searched both optimal solutions of total cost and total time.Other tests also demonstrate similar iteration convergence feature of RGA.Therefore, it is also validated that RGA is effective to solve this stochastic and fuzzy drainage system.

Discussions
Several parameters in this model are adjusted and their sensitivities are discussed to reveal patterns of urban drainage system under extreme rainstorms.

Quantity of Pump Stations.
When quantity  of pump stations is adjusted, both objectives, total cost and total sojourn time, will change.When pump station quantity is 0 through 14, all feasible solutions of the drainage system are compared and shown in Figure 4.
When quantity of pump stations is less than two, none of available solutions can be found out.If pump stations are more than two, total cost increases significantly but total sojourn time of the rainfall decreases dramatically.
When pump stations are more than eight, the curve of total sojourn time declines a little but total cost keeps arising in the same slope, which reveals that a proposal to construct  eight more pump stations in this district is not a good idea considering trade-off between infrastructure investment and drainage capacity under extreme rainstorms.

Diameter of Drainage
Pipes.If diameter of drainage pipe is adjusted, cost of drainage pipes and construction cost as well as its drainage capacity will change.When drainage pipe diameter is less than 600 mm, none of solutions is feasible because stormwater volume at collection points is far more than drainage capacity of drainage pipes.When drainage pipe diameter is more than 600 mm, feasible solutions of total sojourn time and total cost have been found out and shown in Figure 5.With increase of pipe diameter , total sojourn time  decreases and total cost  climbs gradually until its average diameter is not more than 1600 mm.When pipe diameter  is more than 1600 mm, the optimal total sojourn time  remains the same but the optimal total cost  increases still.This result reveals that enlarging drainage pipes will improve urban draining performance, but excessively huge drainage pipe in this district will not decrease stormwater drainage time but will increase drainage system investment.

Precipitation Intensity of Rainstorm.
When precipitation intensity   is adjusted, namely, its Poisson distribution parameter  varies in the model, both objectives are obtained after this model executed many times, and the results are compared in Figure 6.
With increase of precipitation intensity of rainstorm, the optimal total cost  keeps growing and the optimal sojourn time  of stormwater is extended, which means that more investment in urban drainage system is demanded when it is designed to prevent frequent extreme rainstorm.
When precipitation intensity, the distribution parameter , is between 5.5 m 3 /s and 9.5 m 3 /s, the optimal total sojourn time at drainage system increases significantly.When the parameter  is less than 5.5 m 3 /s, the drainage system is able to discharge all stormwater in time, so the curve of optimal total sojourn time is a little flat.
When the parameter  is bigger than 9.5 m 3 /s, for example, an extraordinary rainstorm, this drainage system cannot drain all stormwater in time.Because the total sojourn time is limited by the restriction value   , the total sojourn time  has to be subject to the subjective restrictions and it increases slowly along with the precipitant intensity.It reveals that in a frequent extreme rainstorm district more money should be spent on its drainage system in order to prevent potential urban floods.

Confidence Levels.
If confidence levels of both objective functions and constraint conditions are adjusted, feasible solutions change a lot.Their results are obtained after several executions of the model and shown in Figure 7.
If confidence level is 80%, feasible solutions are from the AB arc.If confidence level is 65%, more feasible solutions are available, and their solution domain is from the CD arc of the curve.When confidence level is higher than 95%, not a feasible solution is available after several trials.When confidence level is 95% or less, more and more feasible solutions can be found out and they all come from the solution domain.
This discussion reveals that confidence levels of objectives and constraint conditions dominate the solution space size to a large extent.When the confidence level is higher than a certain value, not a feasible solution is available, which means the urban drainage system sometimes cannot satisfy all requirements of draining capacities, sojourn times, restricted values, and confidence levels simultaneously.It is suggested that a proper confidence level should be determined based on actual requirements to prevent urban floods caused by extreme rainstorms.

Conclusion
Considering several characteristics in an urban drainage system such as randomness in extreme rainstorm, fuzziness in drainage capacity, and drainage system cost, this optimization model for urban drainage system layout is developed based on a two-stage single-counter queue method.Its two objectives are total sojourn time of stormwater processed by the system and total cost including construction cost and operation cost.Restricted values which reflect draining priority in important areas are introduced to drain stormwater in important areas within limit time.This model can simulate a complex nonlinear problem under stochastic and fuzzy constraints, which represents actual operation of an urban drainage system.
Genetic algorithm is improved and redesigned by embedding with stochastic and fuzzy attributes to solve nondeterministic problems under hybrid constraints, and it is illustrated by a numerical example.It is verified that this model is flexible and effective in solving such a complex nonlinear problem.
The sensitivities of quantity of drainage pump stations, drainage pipe diameter, precipitant intensity of rainstorm, and confidence level on urban drainage system are discussed, respectively, and their results reveal the following: (1) a heavier precipitant intensity will increase total cost of urban drainage system construction and operation and total sojourn time of stormwater disposed by the system; (2) enlarging drainage pipes will increase drainage system cost and its drainage capacity, but total sojourn time of stormwater will decrease dramatically; (3) increasing quantity of pump stations will raise drainage system cost but shorten total sojourn time.But not an optimal solution is available when their quantity is less than a limit.So a minimum pump station quantity should be set in designing an urban drainage system; (4) a higher confidence level of drainage service time can significantly decrease feasible solutions.When confidence level is higher than a limit, not a feasible solution for the drainage system would be found out.
A future research work would be to improve the model features and extend its conditions.The model for drainage pipes will be developed with sophisticated connections and assorted pipes to increase accuracy in modeling an urban drainage system.The simulation algorithm will be upgraded with various distribution functions to represent stochastic and fuzzy characters of extreme rainstorms and drainage capacity.

Disclosure
Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of NSFC.

𝜒 2 𝛼
( −  − 1) = 11.070.Therefore,  2 <  2  ( −  − 1), which means  0 is accepted at such a confidence level.So it is verified that  1 is subject to a Poisson distribution whose parameter  = 7.60.Its mean value  1 = 7.60 and its variance  2 1 = 7.60.Considering some areas are more important in drainage management, eight experts are invited to rank all drainage points.A restriction value   at point  obtained after calculating their rankings by AHP.All restriction values are uncertain, and they are described by their mean values and variances, shown in Table 3.The main parameters of improved genetic algorithm are shown as follows.Population size  = 200, crossover probability   = 0.7, mutation probability   = 0.04, orderbased evaluation function  = 0.05, and maximum iteration number Maxgen = 1000.

Figure 2 :
Figure 2: Optimal layout of the drainage system example.

Figure 4 :
Figure 4: Impact of pump station quantity on total cost and sojourn time.

Figure 6 :
Figure 6: Impact of precipitation intensity on total cost and total sojourn time.

Figure 7 :
Figure 7: Confidence levels and feasible solutions.

Table 2 :
Example of chi-square distribution test for precipitation intensity  1 .

Table 3 :
Restriction value   at each drainage point (mean value and variance) (unit: min).

Table 4 :
Solutions of decision variable   .

Table 6 :
Comparison of average results after 10 operations by RGA, SA-GA, and TS.

Table 7 :
Results after ten operations of Case 1 Total cost including construction cost and operation cost of urban drainage system   : Cost of equipment related to pump station at location  (measured in drainage capacity)   : Cost of construction related to pump station at location  (measured in station)   : Cost of wear and tear of drainage pipe from location  to  (measured in length)   : Length of drainage pipe from locations  to    : Restricted sojourn time for draining stormwater from location    : Precipitation intensity at point  (measured in duration and area) ℎ  : Cost of installation of drainage pipe (diameter ) (measured in length)    : Coefficient converting a drainage pipe coverage area   to its length     : Coverage area of a drainage pipe (diameter ) where groundwater is only collected and conveyed by this pipe   : Duration of rainstorm   : Water capacity of drainage pipe (diameter ) (measured in length) t : Sojourn time of stormwater conveyed from point  to point    : Demand rate of drainage pipe by stormwater at point    : Arrival rate of stormwater at pump station  (measured in time)   : Service capacity of drainage pipe  : Quantity of total pump stations in urban drainage system V  : Flow velocity in drainage pipe  , : Confidence level of objectives and constraints.Variables   : Indicating whether a pump station is located at point  or not: 1, yes; 0, no   : Indicating whether stormwater from point  is disposed by pump station  or not: 1, yes; 0, no   : Draining service ability of pump station at point .