1. Introduction
Drainage systems around the world have experienced an increase in operating pressure in recent years. This is mainly due to the increase in the intensity of rains and urban growth. According to many authors, extreme rains occur with increasing frequency around the world, mainly due to climate change, and are the main factor of flooding in urban basins [
1,
2,
3]. On the other hand, anthropological action has altered the composition of the world atmosphere; one of the most evident is produced by the development of urban space. During the last few decades, cities have experienced a constant process of growth, which has reduced the green areas that surround them, replacing them with highly impermeable surfaces. These factors have led many cities to appreciate the increase in surface runoff, and in many cases, the collapse of their drainage systems [
4,
5,
6]. These problems make it necessary to improve the functioning of the drainage networks to restore security to the cities. Floods in urban areas generate significant economic impacts in cities. Concern increases as cities are increasingly exposed to flood risk [
7,
8]. To face this problem, different options for optimizing drainage networks have been developed.
One of the most prominent approaches is called low impact development (LID). LID approaches have been widely used to minimize the flow of water produced by urban runoff, retaining the water and enhancing its infiltration. This type of approach also attempts to improve water quality by removing pollutants in vegetation and restoring urban ecosystems. Among the most used types of these systems are permeable paving, infiltration trenches, vegetated swales, bio retention basins, and infiltration basins. However, although LIDs are a valid solution to reduce runoff, these approaches do not have great resilience in extreme rain events [
9,
10,
11,
12]. The use of storm tanks in an urban environment to reduce the risk of flooding has been studied and proven as one of the most efficient methods to reduce surface runoff [
13,
14,
15]. Better results have been given by using the combination of storm tank installation and pipe renewal [
16,
17]. The inclusion of hydraulic control elements [
18] appear as an improvement of these works, and turns out to be an alternative that improves the results obtained.
To face the problem of floods, some authors improved the networks by taking as a criterion the avoidance of floods in the study area [
19], while other authors improved the networks based on the flood volume [
16,
20]. Different methodologies have been developed to address this problem. However, the need to find minimal cost designs has led researchers to optimization algorithms. Different evolutionary techniques have been tested for the optimization of drainage networks, highlighting an ant colony optimization algorithm [
21,
22], simulated annealing [
15,
23], harmony search algorithm [
24,
25], and taboo search algorithm [
26]. All of these techniques show good results in different cases studied. However, genetic algorithms that do not require continuity of the objective function stand out in this field due to their robustness, and their use has gained popularity in drainage network optimization work [
10,
16,
17,
27,
28,
29,
30].
Genetic algorithms are stochastic search strategies based on natural selection mechanisms, which involve aspects of biological evolution to solve optimization problems. One characteristic of these types of algorithms is the way that they explore the space solution. While other algorithms follow a single search direction, genetic algorithms perform parallel searches in different directions. This characteristic adds to their ability to explore complex adaptive landscapes, and has made these algorithms a widely used tool in water resources research.
However, one of the problems that most worries researchers is the difficulty that genetic algorithms (GAs) can present in finding solutions close to the global optimum. Kadu et al. [
31] mentions that genetic algorithms are efficient and effective in finding low-cost solutions in drainage system optimization problems. The efficiency and effectiveness depend on several parameters, some that contemplate the parameters of the algorithm and others of the space where the GA seeks the optimal solutions. The problem with the search space (SS) occurs because of the large number of decision variables (DVs) that the algorithm must analyze in a real-life problem. Handling a significant number of DVs causes the SS to grow exponentially. This problem turns into a considerable computational demand to find a satisfactory solution to the problem proposed. Although the computational advance can compensate the time required in this operation, in reality, there are problems in which the SS is so large that it becomes unapproachable. For this reason, the need arises to reduce the SS. This reduction, however, must be done in such a way that the most promising region that contains the best solutions is clearly identified. Maier et al. [
32] mention that a reduction in the size of the SS generally results in its approximation, either because a series of DVs have to be fixed before optimization or because the nature of the interactions between the DVs excludes the effective size reduction. This could potentially exclude the region containing the global optimum, and thus reduce the quality of the solutions found. For this reason, the reduction method must guarantee that the selected region is truly the best, and that the best solutions to the problem are not excluded from it. To solve this situation, different works have been carried out to effectively reduce the SS.
One of the first works was carried out by Schraudolph and Belew [
33]. These authors presented an approach based on dynamic parameter coding to adaptively control the mapping of fixed-length chromosomes to real values, so that, at each iteration, the algorithm searches a smaller SS. Ndiritu and T.M. Daniell [
34] presented a modified GA combining a fine-tuning strategy to reduce the SS with a hill climbing strategy to move to more promising regions. In recent research, Sophocleous, Savić, and Kapelan [
35] presented a model to detect and locate leaks in water distribution networks. The model employs two stages: search space reduction, and leak detection and location. In the first stage, they reduce the number of DVs and the range of values that they can take through an analysis of the characteristics of the network. For the second stage, they used a GA to find the solution to the problem. Simultaneously, Ngamalieu-Nengoue, Iglesias-Rey, and Martínez-Solano presented a methodology to search space reduction (SSR) applied to a drainage network rehabilitation model. The reduction of the SS is based on locating the possible nodes in which storm tanks (STs) will be installed and subsequently identifying the possible pipes that should be renewed. The process is carried out by reducing the number of DVs and the range of values that they can adopt. In a later work Bayas-Jiménez et al. [
18], they used the same methodology to optimize drainage networks. The authors included in the optimization process the use of hydraulic controls that certainly improved the results obtained, but that considerably increased the SS. The proposed methodology continues and complements these works with the aim of improving the efficiency of the optimization process. Specifically, the method reduces the SS using the sectorization criteria to improve calculation times and reduce the computational effort required. To achieve this, an SSR method is applied in each hydraulic sector, decreasing the DVs and defining a final search region. Once the region that is presumed to have the best solutions is delimited, a final optimization is carried out to find the best possible solution. The method is applied to different drainage networks to prove its benefits.
4. Discussion
The methodology presented in this work considers different types of decision variables: pipes, storm tanks, and hydraulic controls. Consequently, there is a large search space. To solve this problem, a process is presented to reduce this search space. The analysis of this discretization has been made considering that the sampling points cover the entire search space and that the process is applicable to any type of network. However, a previous detailed study showed that the characteristics of the network, maximum and minimum diameters installed, and available flood area can shorten the size of the options list for each decision variable by reducing the initial search space and decreasing the definition time of the final search region. This is one of the fields that can be explored in future work. On the other hand, to discretize the continuous decision variables, two spacing values are adopted that generate two lists of options, one called rough and the other called refined. One limitation of the presented method is that the spacing has been established according to the criteria of the authors, and although it is true that covering the search space in the best way has been tried, future works should consider the establishment of this spacing by analyzing the influence of the algorithm parameters.
Another aspect to discuss in this work is the different percentages used as the probability of success when establishing the stopping criteria. The use of a low probability of success (P
e = 20%) in the SSR process aims to eliminate DVs. This elimination assumes that if, with a loose value of P
e, the repeatability of a DV is low, it is assumed that this DV will not be part of the region of best solutions. In the figures (repeatability figures), it can be observed how the repeatability in the most promising DVs increases as the SS decreases, while the DVs that present a low repeatability disappear in subsequent iterations. These facts emphasize the value of the presented method to identify the region with the best solutions within the SS. However, the use of more demanding values of P
e as the search region is reduced is presented as an interesting option to improve the efficiency of the process. This is undoubtedly an aspect that should be analyzed in depth in future work. In this work, the concept of reducing the search space by hydraulic sectors has also been introduced. This way of applying the search space reduction method is presented as an interesting option that allows the optimization model to work in parallel, able to define the search region faster than when analyzing the entire network, although it is true, in the optimization of the Ayurá network, that the hydraulic sectors are not clearly differentiated. Dividing the network into two parts allows us to have a manageable number of DVs by optimizing the network. Applying the SSR process at the top of the network at first and incorporating the selected DVs at the bottom part of the network reduces significantly the computational effort required. In the case of the E-Chico network, two sectors, called HS-1 and HS-2, are clearly differentiated. In this network, it can be seen how the repeatability increases in each iteration until presenting a repeatability of 100% in the selected DVs. Additionally,
Figure 17 compares the dispersion of the solutions of the results when applying the proposed methodology in the studied networks and when optimizing the networks without applying any SSR process. In
Figure 17, it is clearly observed that, when optimizing the complete network, the dispersion of the results is greater, having values of the objective function up to 3.7 times greater than the minimum value, while, when applying the proposed methodology, the dispersion of the results decreases significantly. The curve obtained in the final optimization has a greater slope, and the variation of the results found is reduced by about 1.50 times. In the Ayurá network (
Figure 16), there are similar results; the optimized network without any SSR process has a greater dispersion of results, having objective function values of up to 1.90 times greater than the minimum value found. This shows us that optimizing networks with a large number of DVs increases the dispersion of the solutions. Therefore, applying the SSR process not only reduces calculation time, but also improves efficiency in optimization. These particularities were already observed by Ngamalieu et al. [
16] in a previous work.
On the other hand, the results show that, when optimizing the network by sectors, the values of N and G
max decrease; this is because their value depends on the number of DVs analyzed by the PGA. Therefore, the calculation times and the computational effort are going to be much less than when optimizing the entire network. This shows us that the methodology can be indicated to optimize large networks. In another order of things, the results obtained from N
sim in
Table 5 show that the number of iterations that each simulation requires to find the solution decreases in each application of the SSR process. These results are presented in the two networks studied, so it can be said that the methodology presented improves the efficiency of the optimization. For the final optimization, using a refined optimization and a more demanding Pe, the number of simulations increases. Considering that the objective is to find the best solution in a reduced SS, this increase is fully justified. In the Ayurá network (
Table 6), it is observed that, in each iteration, the iterations decrease, having a N
sim = 3127 in the new scenario until N
sim = 967 in the third iteration. Optimizing the network by sectors has clear benefits. The decrease in the objective function is clearly observed in
Figure 17 when compared to the optimization of the network without applying the SSR process. The figure shows that the largest dispersion of the optimized network without an SSR process is found in the first quartile. On the other hand, it can be observed that the best result of the optimized network without applying the SSR process has a higher value than the worst value of the final optimization of the proposed methodology. In the Ayurá network (
Figure 17), there are similar results, and it is observed that the highest number of results in the final optimization is in the last quartile, which reflects the suitability of the methodology.
The final optimization shows the suitability of applying the optimization methodology with the installation of STs, renewal of pipes, and installation of HCs. The use of a more demanding stopping criterion can guarantee us that the search for the best solutions in the final search region is carried out intensively. For these reasons, this form of optimization should be investigated in greater depth, as it could provide interesting results in terms of reducing calculation times in the face of this type of problem.
Lastly, technological advancement and limited resources have motivated the development of new research strategies to optimize time and economic resources. Algorithms have been consolidated as a valid tool to facilitate this task. Their use in different fields of research has been popularized in recent years. In problems of water resources, GAs have had a relevant use [
32,
39,
40]. GAs have been applied in subjects such as water distribution systems and closely related applications, urban drainage and sewer system applications, water supply and wastewater treatment applications, applications in hydrologic and fluvial modeling, groundwater system applications, groundwater remediation, groundwater monitoring, and evolutionary computation in hydrologic parameter identification. The proposed SSR methodology considers the iterative elimination of DVs from a problem, and can be applied to other types of problems with a similar approach. This may be a contribution of the present work for future developments in the field of optimization through evolutionary strategies.
5. Conclusions
The increase in flood events in different cities of the world makes it necessary to develop methodologies to face them, considering the improvement of the systems with the lowest possible cost. The use of heuristic approaches is a good alternative to solve this type of problem. The methodology based on the optimization of the network considering the replacement of pipes, installation of storm tanks, and inclusion of hydraulic controls has proven to be a valid alternative to solve these types of problems. To improve the efficiency of the optimization model, this work focuses on presenting a methodology for reducing the search space for solutions to improve the working efficiency of the optimization model. The application of this methodology significantly reduces the number of total iterations when compared to the initial scenario, that is, if it will perform the optimization with all of the decision variables.
In the E-Chico network in particular, the SS is reduced from a magnitude of 140 to a magnitude of 12, while, in the Ayurá network, much larger than E-Chico, the SS is reduced from a magnitude of 344 to one of 82. If it is considered that the size of the SS is measured in a logarithmic scale, it can be seen that a significant reduction has been made. These results show that the search space reduction method is valid and advantageous to apply in optimization problems with drainage networks. For the Ayurá network, the improvement of the results can be observed more greatly when applying the proposed method than in the complete network without any previous reduction of the SS. On the other hand, if the results obtained from the E-Chico network are compared with the results obtained in a previous study [
18], it can be seen that the objective function of the problem has been reduced, including the cost of flood damage. Therefore, it can be concluded that by using the method of SSR based on the use of a coarse discretization of the decision variables and a lax stop criterion in the first stage, as well as a refined discretization and a demanding stop criterion in the second stage, the efficiency of the optimization model is improved. The two cases presented have been applied to drainage networks located in Colombia, and thus were expressed in terms of the local Colombian economy, the investment costs in infrastructure, and the costs associated with flood damage. In this way, formulating cost functions in monetary units is very useful for decision makers in the development of a rehabilitation project.
New trends are focused on source control by means of LID techniques. Future works must address the possibility of combining the strategy presented in this work with another focused on the reduction of runoff using components such as green roofs, pervious pavement, or infiltration structures.
Finally, the results obtained show a good solution for a previously defined rain. Different results will be obtained for other design storms. Therefore, there is no single solution to the problem, and the initial approaches to the problem will be made in accordance with design criteria and local regulations.