Optimizing urban flood management: enhancing urban drainage system efficiency under extreme rainfall events

AWS

automatic weather station

ISM

intelligent scheduling model

KP

Korean peninsula

LID

low-impact development

MOPSO

multi-objective particle swarm optimization algorithm

PSO

particle swarm optimization

SWMM

storm water management model

TC

tropical cyclone

TFV

total flood volume

UDS

urban drainage system

WLF

water level fluctuation

The growing vulnerability of urban areas to flooding, which leads to considerable socio-economic damage, can be attributed to the rising frequency and intensity of flood events. These events are often triggered by active tropical cyclone (TC) events (Mahmoud & Gan 2018; Liu et al. 2022). Zhang et al. (2023) assert that over the past four decades, East Asian coastal areas have seen a significant surge in the life span and intensity of TCs. The contribution of TCs to extreme rainfall event and TC-induced precipitation (Lau et al. 2008; Walsh et al. 2019) necessitates a focus on TC-induced extreme rainfall when strategizing flood mitigation. One such strategy pertains to the urban drainage system (UDS), a crucial component of urban infrastructure with notable economic and environmental ramifications (García et al. 2015). However, the TC-induced rainfall overwhelms the design capacity of traditionally designed UDSs, leading to system overflow (Lund et al. 2019). The operation of drainage system pumps, a vital aspect of UDS management, seldom considers climate change and TC influences (Yang et al. 2019). As such, future UDSs must be capable of handling extreme TC-induced rainfall events.

While some researchers propose costly or space-demanding mitigation strategies like low-impact development (LID) measures (Fiori & Volpi 2020) or expanding drainage systems (Zhou et al. 2018), these might be impractical. As an alternative, scheduling strategies that optimize existing infrastructure without additional investment are suggested (Wong & Kerkez 2018; Rathnayake & Anwar 2019). Dynamic scheduling strategies are particularly promising because they can adapt to changes and optimize current infrastructure better than traditional, predetermined operation rules.

The scheduling models available in the literature are diverse, employing various algorithms, variables, and objective functions. While early models usually focus on a single objective like reducing operation costs, current trends point toward intelligent algorithms guiding UDS operations. These smart operations, which apply heuristic and optimization-based methods, make use of genetic algorithms (Jamieson et al. 2007), ant-colony search (Rao & Alvarruiz 2007; Hashemi et al. 2014; Li 2020), fuzzy logic control (Li 2020), and deep deterministic policy gradient algorithm(Ochoa et al. 2019). These methods optimize UDS components such as pumps and detention reservoirs (Chiang et al. 2011; Yazdi 2019; Li et al. 2022).

Multi-objective scheduling strategies are increasingly recognized as essential (Yang et al. 2019). These strategies can better utilize the storage and conveyance capacity of existing facilities by considering multiple objective functions. Still, more existing strategies fail to balance the benefits among objectives due to the use of the multi-objective weighted sum method. Intelligent models for pump scheduling with multiple objective functions have been proposed (Barán et al. 2005; Wang et al. 2009; Zare et al. 2012; Zarei et al. 2022), but they largely ignore the impact of varying moisture sources on UDS operation strategies. To address these gaps, this study aims to design a multi-objective intelligent scheduling model (ISM) that incorporates the effects of varying moisture sources during extreme rainfall events. This model will be developed within a simulation–optimization framework by coupling the storm water management model (SWMM) with the multi-objective particle swarm optimization algorithm (MOPSO).

Rainfall data for this study was collected from an automated weather station (AWS 410) located in the middle of the Dealim3 catchment. A total of 25 years of rainfall data (1998–2022) at 10-min intervals were used. We selected the largest events in the recent flood record to compare the different operating systems of the UDS, using information on rainfall type, duration, statistical data, and typhoon trajectories. The target basin was significantly impacted by major events on 22 September 2010 and 27 July 2011. The estimated rainfall duration and frequency were 240 min and 100 years for the 2010 event and 1,080 min and 100 years for the 2011 event. To calibrate the SWMM model, rainfall data and flow series from the Daelim3 pump station were collected from 20 September 2010, 21:30 to 22 September 2010, 05:50. Similarly, data from 26 July 2011, 15:40 to 29 July 2011, 16:20 were collected for model validation.

The Korean peninsula (KP) is experiencing an increase in TC damage due to the poleward migration of TC tracks (Feng et al. 2021; Wu et al. 2022) and the rise in TC-induced rainfall (Zhang et al. 2023). Urban flooding caused by extreme rainfall events, consisting of TC-induced rainfall and non-TC rainfall, has affected the KP (Wu et al. 2022; Zhang et al. 2023). To examine the performance of the model under different scenarios, TC-induced rainfall and non-TC rainfall are separately input into the model. Therefore, it is necessary to identify typhoon rainfall.

A previous study proposed a method for extracting TC-induced rainfall using gauge-based rainfall data (Kim et al. 2020; Wu et al. 2022). The method identifies TC-induced rainfall based on the ‘TC radius’. If the distance between the TC center and the rain gauge is within the given TC radius, the rainfall occurring before and after that moment is considered TC-induced rainfall. An operation scheme by Yazdi et al. (2016) suggests using different radii for different TC categories. According to Kim et al. (2020), the total duration of rainfall caused by a typhoon in a specific location is represented by the period of the typhoon entering and leaving a 500 km range of the area within a time window of its landfall. Following previous studies, we set the range threshold at 500 km (Son et al. 2017) and the time window at ±1 day (Dhakal & Jain 2020).

To address practical problems, such as easy implementation, unique search mechanism, and fast convergence, PSO has been extensively used (Yang et al. 2009) and applied to urban drainage problems (Wang et al. 2019). However, in optimizing urban drainage and pumping station systems, diverse objectives are often involved. Thus, the MOPSO has been utilized (Yang et al. 2009; Al-Ani & Habibi 2012; Gan et al. 2020).

In this study, the focus is on optimizing the operation of three pump stations and one weir in the study area, which plays a crucial role in controlling the local drainage system. By scheduling the pumps and weir optimally, the storage capacity and conveyance capability of the drainage system can be utilized effectively to prevent urban flooding (Yazdi et al. 2016). The primary objective of UDSs is typically flood prevention and minimization of combined sewer overflow. For this study, the primary objective is to minimize the total amount of flooding. Intelligent algorithms optimize the cooperation between pumping stations and weirs to achieve this objective. The drainage system's maximum capacity is 1.23 × 10⁵ m³. However, due to the complex topology of the drainage system, local flooding occurs through manholes when the inflow exceeds 9 × 10⁴ m³. The SWMM model used in this study reflects flood conditions through overflow at manholes, as it cannot simulate surface inundation dynamics (Zhou et al. 2018). Flooding occurs when surface runoff exceeds the node's capacity, and the total flood volume (TFV) (F₁) is calculated as the sum of runoff exceeding capacity from all manholes experiencing overflow.

To confine the search space and avoid abnormal conditions, constraints are established. For instance, the switch-off water level for a pump must be lower than the switch-on water level to prevent errors in PySWMM. The listed constraints include:

The decision variables in the MOPSO framework dictate the operation of three pumping stations and the associated weir. These variables are influenced by the water level in the front pool, which establishes the switch-on and switch-off thresholds for each pump. With a total of 17 pumps and a weir operating across water levels from 0 to 1.1 m in 0.1 m increments (Goh & Chan 2012), we identify 45 decision variables (Table 1). These include the on/off statuses of the pumps, water level thresholds for activation, and the degrees of weir opening. Such variables are crucial for optimizing operational strategies during various storm scenarios. The intelligent scheduling strategy leverages these variables by analyzing the relationship between water levels and system responses. For example, when the water depth exceeds a specific threshold, MOPSO evaluates the optimal combinations of pump activation and weir openings, ensuring efficient drainage and responsiveness to changing conditions. This dynamic approach ultimately enhances the management of the drainage system.

The SWMM model is invoked in the MOPSO algorithm to obtain corresponding objective function values calculated from the decision variables. The model incorporates the non-linear reservoir method for overland flow routing, Horton's method for calculating infiltration losses, and dynamic wave simulation for conduit flow routing. The decision variables generated by MOPSO serve as scheduling rules for the pump stations and weir in the SWMM inp file. The simulation is conducted using PySWMM (McDonnell et al. 2020), and the resulting objective function values are used in the MOPSO algorithm to determine the velocity and position vectors for the particles in the next iteration.

For the optimal TFV reduction strategy (Table 3), TFV shows a reduction ranging from 10.73 to 57.77%, while WLF is reduced by 26.32–39.54%. Notably, the maximum rainfall intensity of S1 and S3 exceeds the design frequency of this basin including design capacity of facilities, limiting the effectiveness of TFV optimization. However, WLF reduction demonstrates better results. In the case of the longest rainfall duration, TFV is better optimized than WLF. For the optimal strategy with minimum WLF, TFV, and WLF are reduced by 3.11–57.67% and 40.61–54.14%, respectively. While TFV optimization may not be as effective for S1 and S3, a slight reduction in TFV leads to significant improvement in WLF optimization for S2.

Using the TC-induced rainfall identification method described in Section 3.1, we identified a total of 319 TC-induced rainfall events based on 25 years of rainfall data and TC track data. We excluded events with total rainfall below a threshold of 8 mm, as these were considered insufficient to produce flooding (Xie et al. 2023a, 2023b). From the pool of rainfall events that met the minimum threshold of 8 mm, we employed a random sampling method to select nine TC-induced rainfall events. This random sampling method, facilitated by a random number generator, ensured that each event had an equal chance of being selected, thus minimizing any selection bias. The intention behind this approach was to create a representative subset of TC-induced events for further analysis, which is crucial for maintaining the integrity of our study. Detailed information regarding these nine selected events is provided in Table 2 in the previous section.

For the optimal TFV reduction strategy, both objective functions show a reduction for all nine TC-induced rainfall events. TFV reduction ranges from 4.15 to 42.88%. Excluding SAOMAI (where rainfall intensity exceeds the UDS design criteria), TFV reduction rates are above 10%. Additionally, WLF reduction ranges from 24.01 to 58.93%. TFV reduction rates are generally higher than WLF reduction rates.

Similarly, for the optimal WLF reduction strategy, both objective functions are reduced for all nine TC-induced rainfall events. TFV is reduced by 2.32–33.07%, while WLF is reduced by 27.84–65.48%. Compared with the optimal TFV reduction strategy, the reduction in TFV optimization is generally smaller for the optimal WLF reduction, while the improvement in WLF optimization is generally larger.

For WLF reduction, all strategies have a positive impact on the objective function values. In the case of SAOMAI, the intelligent scheduling strategy slightly outperforms the infrastructure upgrade strategies, while the effectiveness of LID strategies is relatively poorer. For KAEMI and MORAKOT, the intelligent scheduling strategy significantly outperforms the other strategies.

For the KAEMI and MORAKOT rainfall events, the intelligent scheduling strategy shows significant reductions in node overflow. For nearly 90% of the peaks, the reduction rate exceeds 50%, greatly mitigating urban flooding in the surrounding area. In the case of KAEMI, some flood peak times are also delayed. Simplified LID strategies do not show significant improvements, as the flood peaks and duration remain similar to the current strategy. Infrastructure upgrade strategies result in an average reduction of flood peaks by 15%, with higher reduction rates as the volumetric extension of the detention reservoir increases. Overall, the intelligent scheduling strategy outperforms the simplified LID strategies and the infrastructure upgrade strategies.

The evidence from the results distinctly affirms the positive performance of the multi-objective ISM in managing both components of extreme rainfall events: TC-induced and non-TC rainfall occurrences. This validates the model's prowess in strategizing the operation of pumping stations and weirs within the study areas, effectively enhancing urban drainage utilization. Thus, through the strategic mitigation of the TFV and WLF, it inherently enhances the stability of the urban drainage infrastructures, thereby reducing the brunt of urban flooding. The reduction in WLF, particularly during high-intensity rainfall events, directly correlates with increased structural stability of the drainage system. The study by Pender & Faulkner (2011) also highlighted the impact of WLFs on the stability of flood defense infrastructure, such as embankments and dams. They warned that rapid WLFs can compromise the structural integrity of these infrastructures, which may ultimately lead to system failure. Studies such as that of Butler et al. (2018) emphasize the need for resilient infrastructure that can withstand both immediate flood impacts and longer-term stressors, such as repeated extreme rainfall events. The reduction in WLF achieved by our model directly contributes to this goal by preventing excessive structural stress on drainage infrastructure components like detention reservoirs and pumping stations.

When scrutinizing the model's functionality with varying rainfall characteristics, it performs markedly well during prolonged periods of rainfall. Nonetheless, under high rainfall intensity – especially when surpassing the UDS's design maximum – the model's effectiveness tends to wane. This implies that the efficiency of the optimized drainage strategy is intrinsically connected to the UDS capacity within the study area.

Each distinct rainfall event allows for the selection of several Pareto optimal solutions, each representing a unique blend of objective function values. This offers the flexibility of choosing the most advantageous regulation strategy, contingent on the priority given to each objective. However, it is crucial to recognize that during certain rainfall scenarios, compromises on the objective do not necessarily result in significant enhancements in another. For example, for rainfall with high maximum intensity, the improvement in WLF optimization is not significant if the TFV optimization efficiency is reduced.

When analyzing the changes in detention reservoir water levels and manhole overflow, the model surpasses the other two strategies. Despite not entirely eradicating flooding at vulnerable nodes, it was able to mitigate flood peaks, indicating the UDS capacity limits its efficiency. When the capacity of the UDS is consistently exceeded, more substantial interventions might become necessary. These could include infrastructure upgrades or the implementation of LID strategies. However, such approaches come with their own set of challenges, including high costs and significant space requirements. Expanding drainage systems in densely populated urban areas may not be feasible due to space constraints, and LID measures often require substantial upfront investment and ongoing maintenance (McGarity 2012). Therefore, while these strategies can be effective, their practicality should be carefully weighed against the specific needs and constraints of the urban environment. Consequently, future research should focus on formulating a cost-effective combination for optimal flood control. A composite approach integrating the model-based strategy and simplified LID strategy/infrastructure upgrade strategy may provide a superior solution.

In summary, the intelligent scheduling strategy, developed through the MOPSO algorithm, has shown significant potential in enhancing UDS management during storm events. By optimizing the operation of 17 pumps and a weir through 45 key decision variables – such as pump statuses, activation thresholds, and weir opening degrees – this approach facilitates effective responses to varying storm conditions. Our revised analysis includes a comparative evaluation of Pareto optimal solutions across different storm scenarios, highlighting the adaptability and robustness of these strategies in addressing specific challenges. This flexibility allows decision-makers to balance competing priorities, whether minimizing TFV or maintaining infrastructure stability, ultimately enhancing the resilience of the drainage system. The study successfully formulated a multi-objective optimization scheduling model that not only reduces flood volume but also improves drainage infrastructure stability. This model-based strategy outperformed local adaptation strategies, positioning it as a viable solution. Future research should explore integrating intelligent scheduling with local adaptation strategies to further enhance drainage capacity during extreme rainfall while optimizing the use of existing facilities.

This research was supported by the National Research Foundation of Korea and the China Science and Technology Exchange Center. This research was partially supported by the KICT Research Program (Project No. 20240127-001). We also appreciate the support of the State Key Laboratory of Water Resources Engineering and Management, Wuhan University.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.