Time-series Multivariate Multistep Traffic Flow Forecasting using Temporal Fusion Transformers

Traffic flow forecasting is a task that is becoming increasingly important in this modern world. People use various navigation systems, including a prediction model that provides the estimated time for reaching a destination and its optimal path. The core of those systems is the traffic flow forecasting model. Although there are many forecasting models currently in use, their performance has proven to be insufficient in many situations. In this experiment, we used Temporal Fusion Transformers (TFT) to predict the average occupancy of real-life traffic from some roadside units in California. Extensive hyperparameter tuning is performed to ensure optimal results for the model. TFT outperformed all the state-of-the-art models with a precision of around 95%. Our implementation of quantile probabilistic forecasting on the data adds another layer of forecasting information and has demonstrated that the actual values mostly lie within the confidence interval range, which makes the model more useful for making navigation decisions in a traffic forecasting system.