TROPICAL CYCLONE INTENSITY PREDICTION USING BP-RNN FROM GPS-DERIVED PRECIPITABLE WATER VAPOR AND SURFACE METEOROLOGICAL DATA

• Main Author: Fahmi Chabibi, Febrylian
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/72746
• Institution: Institut Teknologi Bandung

Tropical cyclones are one of the hydrometeorological disasters that can have catastrophic impacts. Even though the cyclone trajectories prediction model has been developed, the cyclone intensity prediction in the waters of the Indonesian Maritime Continent remains challenging. This study aims to predict the intensity of tropical cyclones using Global Positioning System (GPS) observations and groundbased meteorological sensors. The investigation focuses on the April 2021 Seroja Cyclone that struck Indonesia. The neural network (NN) algorithm with the backpropagation method uses a regression and classification approach. As experimental input, the following time window sizes were utilized: 0 hours, 6 hours, 9 hours, and 12 hours. PWV, ZTD, water vapor partial pressure, temperature, air pressure, PWV rate, and temperature rate are predictor variables. Two years of observational data from the CKUP and CRTE stations were used to train the model. Analysis of variations in training data duration is also conducted regarding the ideal scenario at the previous phase. This study concluded that GPS and meteorological sensors can predict tropical cyclone intensity. A time window can be used to accommodate time series data. Cyclones affect observations inversely with station distance from the cyclone track. At CKUP, the probability of detection (POD) is 89%, and the critical success index (CSI) is 84%. At CRTE, the greatest CSI is 55%, and the POD is 73%. Wind speed forecasts have an RMSE of 1,32 m/s at CKUP and 2,08 at CRTE. Experiments with training data length reveal that longer duration does not necessarily improve prediction model performance.