DEVELOPMENT OF RAINFALL PREDICTION MODEL BASED ON THE WEATHER ELEMENTS AND ATMOSPHERE STABILITY INDEX FOR NON ZOM REGION

• Main Author: SUDARYANTO (NIM. 23814304), EKO
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/26892
• Institution: Institut Teknologi Bandung

Indonesia's geographical location passed by the equator and flanked by 2 oceans and 2 large continents has giving effect to the weather conditions that occur. Region mapping to ZOM (Zone Musim) and Non ZOM (Non Zone Musim) gives an idea that the region in Indonesia has characteristics according to geographical location. For forecaster, making weather forecasts from surface air observation data and upper air observation is not easy, it requires sufficient knowledge and experience, therefore the development of rainfall prediction model for Non ZOM area is needed. <br /> <br /> <br /> The development of the ERNN-BN model (Elman Recurrent Neural Network-Biak Numfor) was made using Artificial Neural Networks with input data of weather element and atmospheric stability index . In the first step used backpropagation to select inputs that have a major impact on the recognition of rainfall classification. Selected 11 inputs from 15 inputs are QFE, TTT, RH, Tmx, Tmn, FFF, EEE, K Index, Sweat Index, TT Index and CAPE Index. The use of RNN (Elman Recurrent Neural Network) is intended to find the output of rainfall forecast with the smallest MSE (Mean Square Error) or MAD (Mean Absolute Deviation). Data for 2010-2016 is used in this study with data from 2010-2014 to training data and data from 2015-2016 to test data. The use of Genetic Algorithms is intended to reduce the inconsistency or fluctuation of ERNN output. <br /> <br /> <br /> For the selected ERNN architecture is 1 hidden layer, 1 neuron, tansig activation function, trainbfg training function, and 12 months training period data, on test data of 2015 has MSE fluctuation value 66.734974 and MAD fluctuation value 4.288539 while in test data 2016 has MSE fluctuation value 105.753973 and MAD fluctuation value 2.173979. Furthermore, Genetic Algorithm is used to optimize the value of weights and bias. For the test data of 2015 and 2016, at the minimum MSE target the MSE fluctuation value is 77% and 97% reduction, while at the minimum MAD target the MAD fluctuation value is 48% and 96% reduction. With this result the systems has created potentially providing additional information on weather prediction in the Biak Numfor area. <br />