Authors: Nugroho Widyanto, Jin-Whan Kim, Umi Narimawati
Abstract:
Currently, there has been improvement in the handling of the pandemic with the availability of vaccines, which has had an impact on improving the economy and starting to return to normal community activities. However, further studies are still needed. So that this pandemic condition can be anticipated in the future, there must be a model that can predict future COVID-19 cases, so that each country can anticipate it by implementing more appropriate policies and handling. In this research, we will propose 2 methods that are popularly used in the case of time-series forecasting involving other factors in the regression model, namely ARIMAX, which is very superior in predicting linear (stationary) models, and bidirectional LSTM, which is a better development than LSTM, which can predict non-linear (non-stationary) models. The hybrid approach used for these two models is expected to provide much better prediction results with a smaller error range, as well as reducing existing variations, compared to using both methods separately. The process of creating a hybrid ARIMAX-Bidirectional LSTM model involves six sub-processes: data transformation using ARIMAX, data normalization (min-max scaling), feature selection using a genetic algorithm, hyper-parameter tuning for ARIMA models and bidirectional LSTM models, prediction models with bidirectional LSTM, and lastly the creation of a hybrid model. The hybrid model was proven to provide much better prediction results than the models run individually. The ARIMAX-Bidirectional LSTM hybrid model has also been proven to be better than the ARIMAX-LSTM hybrid model, as well as other deep learning-based hybrid models. Even though the ARIMAX-Bidirectional LSTM hybrid model can provide very good predictions, its implementation has not been widely carried out, so it is an excellent opportunity to carry out research on its use and utilization. © School of Engineering, Taylor’s University.
Keywords: ARIMAX; Bidirectional LSTM; COVID-19; Deep learning; Forecasting; LSTM; Model hybrid; Time-series
This article can be accessed at https://www.scopus.com/pages/publications/85199480143