Temporal fusion transformers for traffic flow prediction in smart cities

Abstract

Accurate traffic flow prediction plays a critical role in the development of intelligent transportation systems and smart city planning. This thesis explores the application of Temporal Fusion Transformers (TFT), a deep learning architecture designed for interpretable multi-horizon time series forecasting, to model and predict traffic patterns in urban environments. Using real-world traffic datasets composed of spatial-temporal features such as time of day, day of week, road-specific data, and environmental variables, we construct a forecasting pipeline that captures complex dependencies across multiple locations and timeframes. The research includes extensive preprocessing of temporal data, feature engineering, and the implementation of TFT using PyTorch Forecasting. We conduct hyperparameter tuning to optimize model performance and evaluate the model using standard metrics such as Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) on both validation and test datasets. The results demonstrate the model's ability to capture non-linear patterns and temporal correlations, offering interpretable and scalable predictions suitable for real-time traffic management systems. The thesis concludes with visual and statistical analyses of predicted versus actual traffic flows, emphasizing the effectiveness of TFT in smart city applications.