Components of the CPS-based smart flood information system
A 5C architecture—a connection level, conversion level, cyber level, cognition level, and configuration level (Fig. 2)—was used as the guideline to achieve CPS (Lee et al., 2015), one of crucial technologies of industry 4.0. Monostori et al. (2016) mentioned that CPS is one of the most notable advances in the development of computer science and information and communication technologies. Moreover, CPS are broadly applied in manufacturing, decision-making, and energy-related industries (Rho et al., 2016). This study developed an innovative CPS-based smart water system, called DayuSWS, that bridges the gap between the physical world and cyberspace. On-site sensors and a monitoring network record observations in real time and transfer the observations to a big-data platform (connection to conversion level). The field observations are standardized after data preprocessing and are sent to the DayuSWS as inputs for model simulation (conversion to cyber level). The DayuSWS can be integrated with other existing systems or a decision support system (cyber to cognition level) to trigger appropriate active disaster mitigation measures (cognition to configuration levels).
Connection level
The connection level comprises the on-site sensors. Various types of sensors were installed on the basis of the study’s requirements. The connection level is the first level of the DayuSWS that initiates all system functions. Therefore, the information recorded by this level is crucial to trigger the other levels. The sensors were set up after carefully analyzing the locations to ensure that the collected information is suitable. A scenario involving rainfall for 3 h and 150-mm accumulated rain water, with no water outflow from the park, was considered in this study. The simulated results are presented in Fig. 3. The results revealed the west side of the park to have a high risk of flooding, which is consistent with the topology of the area. Because all the vital facilities such as power and water treatment plants are located on the west side of the park, a total of six water gauges were installed on this side, as shown in Fig. 1. Figure 4 is the pictures of all on-site gauges: the gauge locations from left to right and beginning at the top correspond to the numbered locations in Fig. 1. Gauges 5 and 6 were installed to monitor the water levels in pond A and the Hsinchuangzhi drainage, respectively, and are crucial to understanding the water exchange between the park and outside area. Gauge four was installed to monitor the water level in pond B. The rain water collected on the east side is monitored by gauge three and then transferred to pond B. Moreover, gauges one and two were installed to monitor the water collected on the upper side of the park (red rectangle in Fig. 3). A rainfall gauge was installed at location 6 to monitor the park’s rainfall. The real-time observations of all the sensors are wirelessly transferred to the big-data platform at an interval of 10 min.
Cyber level
The connection and conversion levels transform real-time observations to useful information for decision-makers. The information obtained is highly accurate because it is based on on-site observations. However, information with an adequate lead time could not be provided. Therefore, the cyber level is employed to provide forecasts with a longer lead time. In this study, the cyber level was determined to provide flood forecasts with a lead time of 6 h. To obtain the forecasts, a 1D–1D dual drainage model was included, with rainfall forecasts in numerical format serving as the model input. The details are discussed in the following subsections.
Numerical rainfall forecasts
To provide more reliable and timely flood forecasts, rainfall forecast information is crucial. In this study, high-resolution numerical rainfall forecast information from an ensemble numerical weather prediction system in Taiwan was used. The ensemble numerical weather prediction system fabricated by the Taiwan Typhoon and Flood Research Institute (TTFRI) consists of more than 20 ensemble units that are individual numerical weather prediction models with different model configurations. Therefore, a collection of more than 20 numerical rainfall forecasts can be obtained from this system for the same location and time. This system collects worldwide observation data such as temperature, wind, surface pressure, and relative humidity from satellites, atmospheric sounding devices, buoys, aviation routine weather reports, ships, and other available sources. The collected data are then used as boundary or initial conditions for the ensemble units. Then, rainfall forecasts in numerical format are generated according to these real-time weather conditions. The ensemble system aims to provide 72-h numerical rainfall forecasts and generates four runs per day at a 5-km spatial resolution. Hsiao et al. (2012, 2013), Yang et al. (2015), and Wu and Lin (2017) have provided additional details on the ensemble numerical weather prediction system. Moreover, a statistical technique for effectively integrating ensemble forecasts is applied in the proposed system to improve the accuracy of 1- to 24-h rainfall forecasts. Through an artificial neural network (ANN)-based statistical technique, more than 20 ensemble rainfall forecasts are integrated to yield one accurate rainfall forecast. Our preliminary experiments revealed that improved 24-h rainfall forecasts were obtained through the ANN-based integration technique (Wu et al., 2016). Moreover, to further improve the accuracy of short-term rainfall forecasts (i.e., 1- to 6-h rainfall forecasts), real-time radar observation data are included. A radar data assimilation technique, namely the three-dimensional variational data assimilation (3DVar) system, is used. When the 3DVar system is used to rectify the modeled background errors, improved 1- to 6-h rainfall forecasts are obtained. Such forecasts are valuable for warning concerning floods due to convective storms. More details concerning the 3DVar system were provided by Sun et al. (2016). The proposed system uses the aforementioned rainfall forecasts as model inputs to generate flooding forecasts. The data processing flowchart is illustrated in Fig. 5. First, the ensemble mean of all the ensemble units in the TTFRI ensemble numerical weather prediction system is used to yield 1- to 72-h rainfall forecasts. Among the obtained forecasts, 1- to 24-h rainfall forecasts are then replaced by the improved forecasts yielded by the ANN-based integration technique. Finally, the 1- to 6-h rainfall forecasts are replaced by the improved forecasts yielded by the 3DVar system.
Conversion level 1D–1D dual drainage model
This section briefly describes the overall structure of the 1D–1D dual drainage model utilized in the proposed system. The proposed model was retrieved from the Storm Water Management Model (SWMM), which is an open-source model that has been widely adopted in many academic studies (Hsu et al., 2000; Chang et al., 2015). For the initial rainfall-runoff process, the rainfall information is first applied to the RUNOFF module of the SWMM to calculate the inflow discharges of manholes. Next, the information of the inflows at the manholes is directly used as input to the 1D–1D model to simulate the interaction between the below-ground storm sewer flows and the above-ground street flows (Fig. 6). After the 1D–1D model simulations, the surcharges from the manholes are considered as point sources to rapidly identify the high-risk surface inundation areas.
Governing equation of the RUNOFF module
The RUNOFF module of the SWMM conceptualizes a catchment, which pertains to the input rainfall and output discharge, as a nonlinear reservoir. In this representation, the catchment has an inflow due to rainfall and an outflow due to evaporation and infiltration (Fig. 7). According to the conservation of mass, the net change in depth d per unit time t between the inflow and outflow rates of the catchment is displayed in Eq. (1).
$$ \left\{\begin{array}{c}\frac{\partial d}{\partial t}=i-e-f-q,\kern0.75em if\ d>{d}_s\\ {}\frac{\partial d}{\partial t}=i-e-f,\kern0.75em else\end{array}\right. $$
(1)
$$ q=\frac{1.49W{S}^{\frac{1}{2}}}{An}{\left(d-{d}_s\right)}^{5/3} $$
(2)
where i is the rainfall rate, e is the surface evaporation, f is the infiltration rate, q is the runoff rate (refer to Eq. (2)), W is the average channel width, S is the average slope of the catchment, A is the surface area of the catchment, n is a surface roughness coefficient, and d
s
is the storage depth.
Governing equation for the 1D–1D dual drainage model
To obtain the flood routing in the sewer or street, the continuity equation is combined with the momentum equation, as shown in Eq. (3), to produce the average discharge.
$$ \frac{\partial Q}{\partial t}=2U\frac{\partial A}{\partial t}+{U}^2\frac{\partial A}{\partial x}- gA\frac{\partial H}{\partial x}- gA{S}_f $$
(3)
where Q is the discharge in the sewer or street, t is the time, g is the gravitational acceleration, A is the cross-sectional area of the sewer or street, H is the hydraulic head of water in the sewer or street, S
f
is the friction slope from Manning’s equation, and U is the velocity of water in the sewer or street. Details about solving Eq. (3) were provided by Rossman (2015).
Description of the model setup in Huawei Science Park
The Huawei Science Park is an isolated urban drainage area. The 1D–1D dual drainage model in this area, as shown in Fig. 8, includes 88 subcatchments, 142 manholes, and 216 sewers and streets that collect the runoff and transfer the water to the detention ponds, which are located in the southwest corner of the area.