ANALYSIS AND OPTIMIZATION OF DEEP NEURAL NETWORK VARIANTS FOR PREDICTING RAINFALL DATA BASED ON TIME-SERIES MODELS

• Main Author: Kurniawan Dairo Kette, Efraim
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/71842
• Institution: Institut Teknologi Bandung

Reliable and fast forecasting of rainfall intensity can provide significant benefits in climate forecasting research. In this Thesis, the forecasting model is built based on the Deep Neural Network (DNN) algorithm, which has less computational cost compared with Numerical Weather Prediction (NWP) model. The DNN models used are Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Networks (CNN), and CNN-LSTM hybrid. LSTM and GRU models can extract temporal pattern information from sequential data. The CNN model can learn spatial information patterns between variables. Therefore, it is expected the CNN-LSTM model that studies both information patterns can predict rainfall intensity. This Thesis used LSTM, GRU, CNN, and CNN-LSTM models utilizing several observation time steps and possible hyperparameter setups. Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE) is used as a metric to evaluate the error prediction. In this Thesis, the Kernel Shapley Additive Explanations (Kernel-SHAP) method is used to analyze the impact of each observation variable on prediction results from the DNN model. This Thesis obtained two real-world datasets from the Indonesian Meteorology, Climatology and Geophysics Agency (BMKG) and the Climate Hazards Group Infrared Precipitation with Stations (CHIRPS) for five weather observation stations located in Jakarta and West Java. The CNN-LSTM model has the best MAPE and MSE average scores of 8.415% and 0.02617 for predicting rainfall intensity using observational data throughout the year. Meanwhile, the GRU model has the best MAPE and MSE average scores of 9.533% and 0.03203 for predicting rainfall intensity using data in the rainy season. Based on the Kernel-SHAP analysis of each model prediction result, the influential variables with the highest marginal contribution to rainfall intensity prediction are average humidity, then average, maximum, and minimum temperatures.