River Water Level Prediction for Flood Risk Assessment using NARX Neural Network

Abstract

Flood is one of the primary natural disasters in Malaysia and becoming more frequent and on a large scale lately. Not excluded, Sabah encounters repeated floods caused by river overflow. Therefore, an efficient mechanism for flood risk assessment is needed until a more viable solution exists. This paper proposes using the Nonlinear Autoregressive with Exogenous Input (NARX) neural network to model the river water level as an approach for assessing flood risk. The NARX was trained, validated, and tested using the hydrological data obtained at the target areas: Wariu River (Sungai Wariu), Kota Belud, and Padas River (Sungai Padas), Beaufort. Inputs to the NARX are the current and previous water levels at the upstream and downstream rivers and rainfall at the target area. The output is the predicted water level at the downstream river that can be used to assess flood risk. Results show that NARX trained with the Levenberg-Marquardt training algorithm (trainlm) performs best compared to other training algorithms. Results also show that the NARX could predict up to thirty days ahead of water level prediction, with an R2 of 0.75 and above. However, it is more safe to conclude that a reliable prediction for up to five days ahead, with R2 above 0.85 can be obtained.