ANALYSIS OF GENERALIZED SPACE TIME AUTOREGRESSIVE AND ARTIFICIAL NEURAL NETWORK MODEL WITH MINIMUM SPANNING TREE APPROACH OF WEIGHT MATRIX

• Main Author: Widianti, Tamara
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/69239
• Institution: Institut Teknologi Bandung

The time series data used in space-time modeling can be found linear and nonlinear patterns, especially in data with large dimensions or characteristics that are difficult to identify. The space-time modeling that is commonly used is the GSTAR model which is only able to detect linear patterns in the data so that the constructed model is not optimal in representing the data. To detect nonlinear patterns in the data, artificial neural network (ANN) modeling is used. The ability of the ANN model is to be able to learn and obtain information from complex data with the help of a nonlinear function, namely the activation function. In this thesis, we combine the GSTAR model and the ANN model and also develop a weight matrix method needed to build the GSTAR model, namely minimum spanning tree (MST). The MST method show the relationship and the magnitude of the influence between one location and another. Then, combining the GSTAR-ANN model with the MST weight matrix was used to process data on the increase in COVID-19 cases in five provinces on the island of Sumatra. It is very likely that the spread of the COVID-19 virus from one location to another is due to the mobility of people. People moved out from one city to another and from one province to another. Based on the value of mean absolute percentage error (MAPE), it is found that the GSTAR-ANN model gives better results than the GSTAR model in estimation and forecasting, and the use of the MST weight matrix in the two models gets the same good results, with a conventional weight matrix.