Rainfall Forecasting for the Natural Disasters Preparation Using Recurrent Neural Networks

Abstract

Rainfall forecasting is still a concern for researchers considering the increasing uncertainty weather conditions in various tropical regions, including Indonesia. Therefore, a more robust computational model is needed because of uncertainty climate. Deep learning is a method that allows machines to learn time-based data patterns, such as climate data. One technique that is often used for time series data is Recurrent Neural Networks (RNN). However, the selection of climate feature, time segments, and long records of historical data, pre-processing methods, and prediction models largely determines accuracy. This paper proposed Recurrent Neural Networks for weekly rainfall forecasting. It was the rainfall, temperature, and humidity variable each week within a year. The training used LSTM to generalize climate data for the past ten years. Weighting renewal used Stochastic Gradient Descent (SGD) and Adaptive Moment Estimation (Adam). The results showed that the number of datasets and learning rates determine accuracy significantly. So that using data learning last ten years gave more than 96% accuracy of new data and more than 98% of training data.