Using weather factors and google data to predict COVID-19 transmission in Melbourne, Australia: A time-series predictive model

Abstract

Background: Forecast models have been essential in understanding COVID-19 transmission and guiding public health responses throughout the pandemic. This study aims to assess the effect of weather variability and Google data on COVID-19 transmission and develop multivariable time series AutoRegressive Integrated Moving Average (ARIMA) models for improving traditional predictive modelling for informing public health policy.

Methods: COVID-19 case notifications, meteorological factors and Google data were collected over the B.1.617.2 (Delta) outbreak in Melbourne, Australia from August to November 2021. Timeseries cross-correlation (TSCC) was used to evaluate the temporal correlation between weather factors, Google search trends, Google Mobility data and COVID-19 transmission. Multivariable time series ARIMA models were fitted to forecast COVID-19 incidence and Effective Reproductive Number (R _eff ) in the Greater Melbourne region. Five models were fitted to compare and validate predictive models using moving three-day ahead forecasts to test the predictive accuracy for both COVID-19 incidence and R _eff over the Melbourne Delta outbreak.

Results: Case-only ARIMA model resulted in an R squared (R²) value of 0.942, Root Mean Square Error (RMSE) of 141.59, and Mean Absolute Percentage Error (MAPE) of 23.19. The model including transit station mobility (TSM) and maximum temperature (Tmax) had greater predictive accuracy with R² 0.948, RMSE 137.57, and MAPE 21.26.

Conclusion: Multivariable ARIMA modelling for COVID-19 cases and R _eff was useful for predicting epidemic growth, with higher predictive accuracy for models including TSM and Tmax. These results suggest that TSM and Tmax would be useful for further exploration for developing weather-informed early warning models for future COVID-19 outbreaks with potential application for the inclusion of weather and Google data with disease surveillance in developing effective early warning systems for informing public health policy and epidemic response.

Keywords: ARIMA; COVID-19; Forecasting; Internet search queries; Mobility; Weather.

Fig. 1

Map of COVID-19 incidence per capita (per 100,000 population) from 1st August to November 15, 2021, by local government area boundaries in Victoria, Australia with the Greater Melbourne region highlighted. (Image developed using Administrative Boundaries©Geoscape Australia licensed by the Commonwealth of Australia under Creative Commons Attribution 4.0 International licence (CC BY 4.0).).

Fig. 2

Flowchart of ARIMA process for model selection and validation.

Fig. 3

Heat-map correlation for COVID-19 cases and R_eff with weather variables and Google mobility and Google trends, showing negative and positive correlations using Spearman's rank correlation.

Fig. 4

Daily distribution of COVID-19 cases, Effective reproductive number, Transit station mobility and maximum temperature for Melbourne, 1st August – November 15, 2021.

Fig. 5

Scatter plots of COVID-19 incidence with Tmax (top left) and TSM (top right) and Reff with Tmax (bottom left) and TSM (bottom right) showing the bivariate relationship and linear trends with maximum temperature (left) and transit station mobility (right).

Fig. 6

Time series cross-correlation for COVID-19 cases with Tmax (top left) and Transit station mobility (top right) with and R_eff with T_max (bottom left) and TSM (bottom right) showing 10-day lag for the study period August 1, 2021–November 15, 2021. Confidence interval (95%) is indicated by dotted lines (X-axis lag value, Y-axis CCF value).

Fig. 7

Prediction values for best fit ARIMA models Model 4: COVID-19 cases with maximum temperature and transit station mobility and Model 5: R_eff compared with observed values in Greater Melbourne. (X-axis: Date (daily) Y-axis: COVID-19 cases or R_eff with 95% confidence interval: upper confidence limit (UCL) and lower confidence limit (LCL).