**1. Introduction**

Floods are one of the most threatening hazards to civilian safety and infrastructures, causing damages and losses over the world [1]. Especially in densely populated urban areas, urbanization, aging of drainage systems and climate change contribute to growing flood risk in many countries. Ione important mitigation measure is the prediction of future flood occurrences. Real-time flood forecast with sufficient lead time can boost the use of preventive measures for flood mitigation. Such measures can minimize the threats to communities and individuals at risk of flooding [2].

Various types of hydrology and hydraulic models are available for flood forecasting [3]. From the different modeling types, these can be classified into rainfall-runoff models, one-dimensional (1D) model, two-dimensional (2D) model, and coupled 1D–1D model and 1D–2D model. From the forecast application perspective, 2D and 1D–2D models are able to provide directly a spatial surface flood representation, which is essential for flood damage estimation. 1D–1D models, on the other hand, relies heavily on GIS pre- and post-treatments. Most 2D hydrodynamic models are computationally expensive [4]. Even with the help of up-to-date computational techniques for 2D simulations, the computational capability is still inadequate for a real-time forecast [5].

Data-driven is a fast-growing alternative to hydrodynamic models due to the development of computing technologies in recent years. Data-driven models ignore the physical background of a problem and rather explore the relation between the input and output data [6]. For short or long-term flood forecasts, different data-driven models have been implemented, such as neuro-fuzzy [7], support vector machine [8], support vector regression [9,10], Bayesian linear regression methods [11] and artificial neural network (ANN) [12]. Among them, artificial neural networks (ANN) can be an effective tool for flood modeling, if it is properly applied, overcoming pitfalls as over-fitting/under-fitting with sufficient and representative data for model training [13]. Dawson and Wilby applied ANN to conventional hydrological models in flood-inclined catchments in the UK in 1998 [14]. After that, numerous examinations about flood forecasts in catchment scales emerged [2], [15]. Sit and Demir [16] integrated the river network spatial information to improve the forecast accuracy in Iowa. Bustami et al. applied the backpropagation ANN model for forecasting water levels at gauging stations [17]. ANN showed its great potential on the short-time forecast of extreme water levels with a comparable accuracy to the physical model with a far less computational cost [18]. Simon Berkhahn et al. applied an ANN for two-dimensional (2D) urban pluvial inundation extent forecast [19]. Lin et al. applied backpropagation networks for maximum flood inundation extent prediction and achieved a comparable accuracy to the hydraulic model [20].

The objective of this work is to develop a multistep flood forecast method for urban areas at a fine spatial resolution of 4 m by 4 m. Unlike Lin et al. [20], in this study, an ANN-based framework is proposed for performing multistep forecasts for 1–5 h. To the author's knowledge, only the works of Chang et al. [2] and Shen and Chang [21] were able to produce multistep flood forecasting maps. The novelty herein is the forecast at a finer resolution of 4 m × 4 m, suitable for urban flood forecast. Our hypothesis is that the ANN forecast model can provide comparably accurate flood extents as hydraulic models, but with a much shorter time of several seconds. As the hydrodynamics-based model runs usually takes several hours, the reduction in forecast time of ANN enables more flood mitigation measures to become effective. In Section 2, we introduce the resilient backpropagation artificial neural network structure and the validation of the model. Section 3 describes the study area and the data preparation of the event database. Section 4 shows the results of flood forecasts of the first intervals for synthetic and historical flood events and the real-time forecast for historical flood events. Sections 5 and 6 are the discussion and conclusion of this work.

#### **2. Methods**
