Causal Framework for Traffic Flow Forecasting

This document is an extended abstract submitted for presentation at the TRC-30 conference, focusing on a causal, theory-informed framework for traffic flow forecasting. The authors propose a novel methodology that integrates Granger causality-inspired feature selection with a multitask Long Short-Term Memory (LSTM) neural network to predict two traffic variables simultaneously. The framework aims to address the challenges associated with deep learning models in traffic forecasting, such as data requirements and limited explainability. A custom loss function, termed Traffic Flow Theory-Informed (TFTI) loss, is introduced to enhance model performance by incorporating theoretical traffic flow principles. The methodology is tested using data from the Traffic Management Center of the Region of Attica, which includes measurements from loop detectors in Athens. Results indicate that the proposed approach improves the accuracy and trustworthiness of traffic forecasts, demonstrating its potential applicability to other urban road networks.