ANN for Urban Flood Flow Modelling Using Real Time Data

Abstract

Floods are frequent and devastating natural disaster that can result in significant damage to property and loss of human life. Accurate flood flow modelling is essential for predicting the effects of floods and developing effective flood management plans. Artificial Neural Networks (ANNs) offer many advantages over the conventional method of flood flow modelling based on Saint–Venant equation which is based on many assumptions such as channel geometry and friction while the ANNs have ability to model the relationship without relying on explicit mathematical equations and assumption. For assessing the impact of climate change, accurate forecasting of rainfall is critical. ANNs have proven to be well-suited to address the varied and dynamic nature of flood data due to their ability to learn from vast datasets. Therefore, there is a need to efficaciously use ANNs for flood flow modelling due to their ability to capture complex correlations between input and output data. Keeping the above description in mind, this present study is undertaken to apply feed forward neural network (FFNN) and recurrent neural network (RNNs) to predict the discharge of the river in urban area using meteorological data ensuring minimal computational errors. The study is conducted on the Annapurna ghat gauging station of Barak River in the Cachar district near railway station of Silchar, using data collected from the water resource department of Cachar district, Silchar. The results demonstrate that both the feed forward neural network (FFNN) and recurrent neural network (RNN) models provide better accuracy than traditional regression models such as polynomial, logarithmic, and exponential regression models. Three-layer ANN structure is proposed having 2 hidden layer and single output layer having 3 nodes in input layer and 25 and 5 nodes in each hidden layers respectively with one node in output layer which is denoted as 3-25-5-1. The models’ performance was assessed using RMSE (root mean square error) and R2 (coefficient of determination). The findings indicate that using precipitation, maximum and minimum monthly head as input to ANNs, feed forward neural network (with R2 = 0.9428 and RMSE = 0.1113 m3 s−1) and recurrent neural network (with R2 = 0.8858 and RMSE = 0.1347 m3 s−1) are the most accurate models. For large and complex datasets, RNNs are best suited while for small datasets, feed forward networks provide better and accurate forecasting. The use of ANNs in urban flood flow modelling is therefore recommended for better accuracy and efficiency which is critical in the face of the increasing frequency and severity of floods caused by climate change. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.