A Robustness Analysis of Different Nonlinear Autoregressive Networks Using Monte Carlo Simulations for Predicting High Fluctuation Rainfall

Abstract

In this study, the main objective is to carry out the robustness analysis of an artificial intelligence (AI) approach, namely nonlinear autoregressive neural networks (NAR) using Monte Carlo simulations for predicting the high fluctuation rainfall. Various algorithms of the NAR including Levenberg–Marquardt (LM), Bayesian regularization (BR) and scaled conjugate gradient (SCG) were developed. Statistical criteria, namely coefficient of determination (R2), root mean squared error (RMSE) and mean absolute error (MAE), were used to quantify the impact of fluctuations on the prediction output. Results showed that SCG algorithm was not sufficiently robust, while LM and BR methods exposed a strong capability in forecasting daily rainfall. In addition, prediction using BR was slightly better than LM, especially in terms of standard deviation of R2, RMSE and MAE distributions over 500 Monte Carlo realizations.