Temporal fusion transformer-based forecasting of COVID-19 infection trends using environmental indicators

The COVID-19 pandemic highlighted the severe threat infectious diseases pose to global public health and the urgent need for effective and robust surveillance systems. In response, predictive modeling and time series analysis have gained prominence for forecasting infection trends and supporting public health interventions. This study developed and evaluated a multivariate time series forecasting model based on a temporal fusion transformer (TFT), which effectively incorporates both static and time-varying input variables to predict COVID-19 case dynamics based on environmental factors such as wastewater quality, air quality, and weather conditions. It was developed and tested using a dataset consisting of district-level confirmed COVID-19 cases and environmental variables collected between February 2020 and May 2022. The results revealed that incorporating environmental variables improved the forecasting performance of the proposed TFT-based model by 17 %. The model also demonstrated the ability to capture the nonstationary and time-dependent nature of COVID-19 data, achieving an R² value of 0.962. However, because the number of cases was underestimated in the testing dataset, direct case number predictions proved less reliable. Subsequent indirect estimation based on the proportional change in case numbers yielded a higher predictive accuracy (R² = 0.984). These findings illustrate the value of deep learning approaches for epidemiological forecasting and highlight their potential to support timely public responses and optimize medical resource allocation.