Performance Enhancement Model for Rainfall Forecasting Utilizing Integrated Wavelet-Convolutional Neural Network

The core objective of this study is to carry out rainfall forecasting over the Langat River Basin through the integration of wavelet transform (WT) and convolutional neural network (CNN). The proposed method involves using CNN for feature extraction to efficiently learn from the raw rainfall dataset. With the aid of deep architecture, a highly abstracted representation of the inputs time series with a high level of interpretation is formed at each subsequent CNN layer. The use of WT in forecasting the rainfall time series is by preprocessing the raw rainfall dataset into a set of decomposed wavelet components as inputs for the CNN model using discrete wavelet transform (DWT). The conditions for discretizing the raw input through DWT are discussed, along with the criteria to be used. Daily datasets, ranging from January 2002 to December 2017, were used. The results showed that the proposed model could satisfactorily capture patterns of the rainfall time series, for both monthly rainfalls forecasting or daily rainfall forecasting. Three performance indices were used to evaluate the model accuracy: RMSE, RSR, and MAE. These statistical indices have a range of value from 0 to a finite value that depends on the scale of the number used. In general, a lower value is better than a higher one.