Storm Water Management Model (SWMM) developed by the United States Environmental Protection Agency (EPA) has been widely applied throughout the world for analysis associated with stormwater runoff, combined sewers, and other drainage facilities. To appropriately manage the runoff in urban areas, an integrated system including the simulations of sewer flow, street flow, and regional channel flow, called the 1D/1D SWMM model, was advocated to be performed. Nevertheless, the execution efficiency of this integrated system still needs to be promoted to meet the demand for real-time forecasting of urban floods. The objective of this study is to seek an alternative for predicting water levels both in the sewer system and on the streets within an urban district during rainstorms by utilizing a dynamic neuron network model. To strengthen the physical structure of the artificial intelligence (AI) model and simultaneously make up for the lack of measured data, simulation results of the 1D/1D SWMM model are provided as labels for the training of the proposed model. The novelty of this research is to propose a new methodology to effectively train the AI model for predicting the spatial distributions of depths based on the hydrologic conditions, geomorphologic properties, as well as the network relation of the drainage system. A two-stage training procedure is proposed in this study to consider more possible inundation conditions in both sewer and street (open channel) drainage networks. The research findings show that the proposed methodology is capable of reaching satisfactory accuracy and assisting the numerical-based SWMM model for real-time warning of drainage systems in the urban district.

The design flood can be determined by transposing the design storm observed in the region surrounding the project and converting the storm to flow by use of a simple rainfall-runoff relation or a watershed runoff model. In this study, a design hyetograph in a double-triangular shape of 48 h duration was derived from rainfall intensity–duration–frequency curve to represent the time distribution of severe typhoon rainstorms. The design floods generated by performing a rainfall-runoff model using design hyetographs as inputs were verified using the flow records from 17 watersheds in Taiwan. The result indicated that the generated design discharge using the proposed hyetograph for specified return periods were close to the discharge obtained from frequency analysis using flow records. The proposed design hyetograph can be applied to areas that are usually threatened by typhoon rainstorms.

Several algorithms for flow-path determination have been proposed to pursue a more realistic simulation for flow transmission in watersheds. However, the diversity of existing algorithms may lead to significant differences in the extraction of geomorphological properties. Therefore, preliminary knowledge about these approaches is necessary. In this study, the five most widely used methods – the D8, D , MD , D4, and MD8 algorithms – for flow-direction determination were implemented to investigate watershed extent, water mass contribution, and the track of the longest watercourse. Distinctions of these properties should affect subsequent applications of hydrological analyses, such as the time of concentration calculation and peak discharge estimation. For an example case in Taiwan, the watershed extent and the longest watercourse obtained from the multiple-direction (MD8) algorithm could reach 1.2 and 1.8 times, respectively, that derived from the single-direction (D8) algorithm; moreover, the multiple-direction (MD8) algorithm manifested the weakest ability of collecting water mass from the watershed. A simple index was developed in this study to rank the effects of flow dispersion and flow concentration among the different flow-direction algorithms.