Analysis on Dimensionality Reduction Techniques for Sub-Seasonal to Seasonal Rainfall Prediction

Abstract

Sub-seasonal to seasonal ($S$2$S$) weather prediction refers to a prediction of environmental conditions made in the range of 2 weeks to 12 months. The $S$2$S$ products are based on the output of global climate models, and can be developed further through various statistical downscaling approaches. The training of the statistical models requires relatively high computational resources due to the large size and dimension of Global Climate Model (GCM) output data. This research analyzes the use of several dimensionality reduction techniques that can be used to reduce the dimension of the GCM output data. The compared techniques are one-dimensional Principal Component Analysis (1D-PCA), one-dimensional wavelet decomposition (1D-WD) and two-dimensional wavelet decomposition (2D-WD). Backpropagation algorithm is utilized to train the neural network model using the dimension-reduced GCM data. Simulation results revealed that the 2D-WD model has a relatively consistent performance compared to the other models and has the lowest training time among others. This method has the potential to produce a prediction model with good accuracy and with reasonably low computational cost.