Comparation of Elman neural network, long short-term memory, and gated recurrent unit in predicting dengue hemorrhagic fever at DKI Jakarta

Abstract

Dengue hemorrhagic fever (DHF) is a seasonal disease that has quickly spread throughout the world in the past few years. It is caused by the dengue virus that is transmitted by a biological vector in the form of mosquitoes. In DKI Jakarta, the capital of Indonesia, there were 970 DHF cases at the beginning of 2020, thus placing Jakarta in the red zone for the spread of DHF. Besides, DHF can emerge due to various weather and climate factors such as humidity, rainfall, and temperature. With a significant increase in the number of potentially fatal DHF cases in DKI Jakarta, preventing DHF outbreaks is recommended. This research aims to predict the number of DHF cases in DKI Jakarta using weather and DHF case data. Three machine learning models were employed to predict DHF case numbers: Elman neural network (ENN), long short-term memory (LSTM), and gated recurrent unit (GRU). ENNs have a simplified recurrent neural network (RNN) structure, LSTM is a modified RNN with long-range memory and an activation function for deciding which information should be retained or discarded, and GRUs are modified RNNs that are slightly simpler than LSTMs. These methods were implemented in three different data sets as follows: one with 90% training data and 10% testing data, another with 80% training data and 20% testing data, and finally, one with 70% training data and 30% testing data. A grid search is used to determine the best hyperparameter from all three methods. This will determine the best model for predicting the number of DHF cases in DKI Jakarta (excluding Kepulauan Seribu Regency) according to the data used, RMSE values, and the simulation results. Based on these criteria, LSTM is better suited to predicting the number of DHF cases than ENN or GRU in almost every district in DKI Jakarta.