Improvement of Two-Dimensional Flow-Depth Prediction Based on Neural Network Models By Preprocessing Hydrological and Geomorphological Data

Abstract

The stability and efficiency of a rainfall-runoff model are of concern for establishing a flood early warning system. To tackle any problems associated with the numerical instability or computational cost of conducting a real-time runoff prediction, the neural network (NN) method has emerged as an alternative to calculate the overland-flow depths in a watershed. Therefore, instead of developing a new algorithm of machine learning to improve the predicted accuracy, this study focuses on thoroughly exploring the influence of input data that are highly related to the flow responses in space, and then establishing a procedure to process all the input data for the NN training. The novelty of this study is as follows: (1) To improve the overall accuracy of the 2D flood prediction, geomorphological factors, such as the hydrologic length (L), the flow accumulation value (FAV), and the bed slope (S) at the location of each element extracted from the topographic dataset were considered together and were classified into multiple zones for separate trainings. (2) An optimal length of the effective rainfall condition (T-o) was proposed by conducting a correlation analysis to determine the most informative precipitation data. In this study, the outcomes of four types of NN models were examined and compared with one another. The results show that the simplest structure of the NN methods could achieve satisfactory predictions of flow depth, as long as the approaches of data preprocessing and model training proposed in this study were implemented.