Comparative Analysis Of Neural Network Model Selection And Data Transformation For Rainfall Forecasting

Abstract

The selection of input models in neural networks significantly influences predictive accuracy in time series forecasting. This study evaluates different input models for neural networks in rainfall prediction using data from the Wonorejo Reservoir, Surabaya. The neural network inputs are determined based on significant lags identified through the Partial Autocorrelation Function (PACF) and ARIMA models. Simulation results indicate that the best Feed Forward Neural Network (FFNN) model utilizes PACF-derived input lags and is trained using the Rprop+ algorithm with a logistic activation function. Meanwhile, the optimal Deep Learning Neural Network (DLNN) model employs the Rprop- algorithm with a logistic activation function. The best-performing model for rainfall forecasting, based on the lowest Root Mean Squared Error of Prediction (RMSEP), is the FFNN model with an (8,4,1) architecture. To further refine the model, we applied a stepwise selection process to eliminate non-significant lag inputs. However, results show that this optimization had no substantial impact, as RMSEP increased after the stepwise procedure.