Evaluating Rainfall Predictions using Statistical Decomposition and Neural Network–Based Approaches

Abstract

The ability to accurately predict impacts of rainfall is essential for managing our water supply, developing agricultural plans and protecting ourselves from natural disasters. For the purpose of this study, this research compares statistical and neural network-based methods used in the prediction of rainfall, particularly Singular Spectrum Analysis (SSA), Artificial Neural Network (ANN) and Long Short-Term Memory (LSTM). According to those models, time series can be decomposed into trend, seasonal and noise components during SSA approaches. Meanwhile, other deep learning models offer solutions that utilize historical data to model nonlinear rainfall behavior and account for random events. Data collected from Mersing on daily rainfall are analysed for statistical properties throughout this study and used to develop and test forecasting models. In this regard, metrics can be analysed to determine the performance of models such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of correlation (R) and Nash-Sutcliffe Efficiency (NSE). The results indicate that SSA performed the best reconstruction with MAE=4.1931, RMSE=6.8078, R=0.9362, and NSE=0.8727, however both recurrent and vector SSA had poor forecasting scenarios with negative NSE values. Based on training results, LSTM model showed MAE=4.0391, RMSE=10.5818, R=0.7495 and NSE=0.5307 which indicates that the model has a strong capacity to learn. Although the ANN model outperformed the LSTM model marginally in the test, it achieved a MAE of 9.9117 and NSE of -0.2332 which gives it a marginal edge over the LSTM model. It is suggested that rather than concentrating solely on statistical or neural models, both structured and chaotic rainfall behaviours in Malaysian coastal areas can be better understood by integrating SSA with deep learning techniques for a more effective understanding of the phenomenon. A combined approach can provide valuable insights into the implementation of decomposition-based and data-driven forecasting methods specifically for moist climates. Thus, the findings established the foundation for further studies on developing improved and enhance hybrid rainfall forecasting systems.