Big data analytics for smart transportation

Abstract

As the cost of storage decreases, more data is being collected and if these data can provide us with insights, we will be able to make more well-informed decisions. Therefore, being able to draw insights from traffic data will enable road users to better plan their routes while the relevant authorities in charge of road planning will be better able to allocate land for transport more efficiently. In this report, we will look at two different approaches available for traffic flow forecasting: parametric and non-parametric. All the models will be tested using a traffic flow dataset which is converted from raw GPS coordinates of taxis in Singapore. Both approaches will be evaluated against the Historical Average Model, which is the selected baseline model, using a set of evaluation metrics which are root mean squared error (RMSE), mean absolute error (MAE), accuracy, variance and R2 score. For the parametric approach, models such as the AutoRegressive Integrated Moving Average (ARIMA) was tested. For the non-parametric approach, models such as the Long Short Term Memory (LSTM) that capture only the temporal features of the dataset as well as models such as the Temporal Graph Convolutional Network (T-GCN) that capture both the temporal and spatial features of the dataset were tested. Therefore, the focus of this project is to determine which approach is more suitable for traffic flow forecasting and whether capturing both temporal and spatial features of the dataset will improve the model’s performance based on the set of evaluation metrics mentioned above.