Dengue case forecasting using multi-step deep learning models with attention layers

Abstract

Dengue is a viral infection that is transmitted from mosquitoes to people. It is more common in regions with tropical and subtropical climates. Accurate dengue forecasting is important to make the right decisions on time. In this sense, in this study, deep learning models with attention mechanisms such as long short-term memory (LSTM), bidirectional LSTM (BiLSTM), gated recurrent unit (GRU), and bidirectional GRU (BiGRU) were implemented, and to improve the accuracy of model results they were linearly interpolated. According to the results, in most cases, linear interpolation improved the implemented deep learning models with attention mechanisms in terms of mean squared error (RMSE), mean absolute percentage error (MAPE) and R2. For one-step predictions, improvements occurred between 0.08% and 0.13%, for two-step predictions between 8.55% and 22.81%, for three-step predictions between 0.26% and 23.88%, for four-steps between 0.15% and 4.79%, and between 0.11% and 0.19% for five-step predictions. Based on the obtained results, it is possible to experiment with other types of interpolations such as polynomial, spline, and inverse distance weighting (IDW).