Forecasting of Individuals' Movement in Society After Corona Pandemic Using RNNs with LSTM and GRUs

ABSTRACT

The health crisis that attributed to the quick spread of the COVID-19 has impacted the globe negatively in terms of economy, education, and transport and led to the global lockdown. The risk of the COVID-19 infection has been increased due to a lack of a successful cure for the disease. Thus, social distancing is considered the most appropriate precaution measure to control the viral spread throughout the world. In this study, a model was proposed for deep learning capable of predicting the movement of people in the pandemic in the short term (one day) to take precautions and control the COVID-19 infection. The proposed model consists of four phases data collection, pre-processing phase, prediction stage, and evaluation and Comparison phase. The dataset is obtained from 428 mobility reports, collected based on data from users that have been selected for their Google Account location history for a country such as Iraq for 428 days. A deep learning algorithm such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and hybrid model (GRU&LSTM) is applied to pre-processed data to predict the movement of people. They are compared using statistical measures: Mean absolute error (MAE) and root mean square error (RMSE) for performance measurement of these machine learning algorithms. The results of the GRU are the sum of MAE 0.4277 and sum of RMSE 0.6470 for predict person path and movement with training time equal to 33.189 sec, while the results of the hybrid model are the sum of MAE 0.4355 and sum of RMSE 0.6563 for prediction and the training time equal to 53.144 sec, and the results of the LSTM are the sum of MAE 0.4395 and sum of RMSE 0.6612 for prediction and the training time equal to 100.752 sec. These statistical measurement values indicate proposed model GRU outperformed all other models, it showed a solid performance to predict person path and movement in coronavirus pandemic and took little time to train compared to other algorithms, while the hybrid algorithm showed good performance and a short period in training compared with the LSTM model. © 2021 IEEE.