Convolutional Neural Network with Long Short-Term Memory on Time-Series Rainfall Data Prediction Model Efraim Kurniawan Dairo Kette (a*), Finny Oktariani (b)
a) Department of Computational Science, Bandung Institute of Technology, Indonesia
20920301[at]mahasiswa.itb.ac.id
b) Combinatorial Mathematics Research Group, FMIPA, Bandung Institute of Technology, Jalan Ganesha 10, Bandung 40132, Indonesia
Abstract
Understanding future weather conditions are essential in planning and handling the negative impacts that may arise. The Numerical Weather Prediction (NWP) is a model often used for weather forecasts. However, it usually has high computational costs. In this study, we use Convolutional Neural Network (CNN) combined with Long Short-Term Memory (LSTM) Network model to predict rainfall intensity based on time series data. We use two datasets at five observation stations of the Indonesian Agency for Meteorological, Climatological, and Geophysics (BMKG), along with Climate Hazards Group Infrared Precipitation with Stations (CHIRPS). The CNN model can retrieve spatial information from weather measurement variables, and LSTM maintains temporal patterns information from sequential data to improve the accuracy of the prediction model. We use several observational time steps to emulate short term and long-term prediction. The results obtained using optimal model hyperparameter show that CNN combination with the LSTM model obtains an 8.410% MAPE score whereas the CNN model give MAPE score of 8.688%. Based on the results, we consider that the CNN-LSTM model can improve the rainfall intensity prediction in different weather observation regions.
