Flood Forecasting Using Non-linear Autoregressive Exogenous Neural Networks with Radial Basis Function

Abstract

River stage and discharge monitoring plays a significant role in making sound forecasts of floods, early warning, and preventing risk posed by disasters, especially when rain falls heavily in the areas of river catchment. These variations in the water level are difficult to model and the time series models which are commonly used to recapitulate such a behavior are not able to reflect the complexity. The paper is a discussion on the importance of enhancing the accuracy of flood forecasting by combining two forecasting models, that is, Radial Basis Function (RBF) and Nonlinear Autoregressive Exogenous (NARX) with simple averaging and weighted averaging techniques, and comparing the performance of the two in terms of their respective impact. With the help of a genetic algorithm, the most suitable combination of weights is sought, and the contribution of each of the individual models is adjusted. Therefore, models used singly and intelligent committee machine learning model (NARX, RBF, ENS-AVG, and ENS-GA) are used to generate river level forecasts of the Selangor River at the lead times of 1, 3, and 6 hours. Accuracy of the models is measured using root mean square error (RMSE), mean absolute error (MAE) and the coefficient of determination (R-squared). These results suggest that the weighted averaging (ENS-GA) offers better results in most of the assessment measurements in comparison to the simple averaging. The magnitude of difference between the values of RMSE between the 1-hour forecast as a result of substituting simple averaging with weighted averaging by the use of GA is approximately -98.39% and the 3-hour forecast is approximately 78.57% and 6-hour forecast is approximately -92.86%. The proposed strategy implies the establishment of a consistent and correct time-series forecasting model by integrating a set of forecasting tools and streamlining their work, which will then have a beneficial effect on the quality of flood forecasts.