Multivariate Long and Short Term LSTM-Based Network for Traffic Forecasting under Interference: Experiments during COVID-19

Due to COVID-19, work-from-home policy and travel restrictions were taken to decelerate the virus spreading. While these policies successfully eliminated the transmission of COVID-19, original traffic patterns have been completely disrupted, including considerable reductions in travel time and vehicle miles traveled. The impacted traffic patterns from the unexpected event brings challenges to the U.S. Department of Transportation and transportation planners. With fluctuated traffic conditions, it is difficult for transportation agencies to learn representative traffic patterns from short-term historical data. Therefore, we proposed a multivariate long and short-term LSTM-based model (var LS-LSTM) for network-wide traffic forecasting under interference. We considered multiple spatial and temporal features to evaluate network-wide traffic performance and forecast the influenced travel behaviors. Multi-dimensional spatial-temporal features were fused into long-term and short-term historical matrices and fed into the model, which enabled the model to accommodate intervention from unexpected events. Thorough experiments were conducted using loop detector data in the Greater Seattle Area from 2020 to early 2021 and achieved reliable prediction performance in both robustness as well as accuracy. The proposed model showed competitiveness against other state-of-art algorithms in all experiment time frames, from pre-COVID-19 to COVID-19-relieving period. This study would benefit government agencies and the general public in making sustainable policies and future resilience plans for post-pandemic smart cities.