GNN-Based Traffic Flow Forecasting and Its Future Scope in Nepal

Abstract

In modern urban environments, traffic congestion poses a significant challenge for intelligent transportation systems, which demand accurate and scalable traffic flow forecasting solutions. Conventional time series approaches fail to capture the spatial dependencies inherent in road networks, which motivates the use of graph-structured models. This paper presents a comparative evaluation of three approaches: Linear Regression, Long Short-Term Memory (LSTM) networks, and Spatio-Temporal Graph Convolutional Network (STGCN) for short-term traffic speed forecasting on the METR-LA benchmark dataset, which consists of 207 highway sensors. The models are evaluated using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) at a 15-minute prediction horizon. As a result, STGCN achieves an MAE of 2.18 mph and RMSE of 3.75 mph, outperforming the LSTM baseline by 29.7 % on MAE and Linear Regression by 18.4% on MAE, showing that jointly modeling spatial road network topology and temporal traffic dynamics significantly improves forecasting performance. Finally, we analyze the applicability of graph-based traffic forecasting to the rapidly urbanizing road networks of Nepal, where the lack of sensor infrastructure and data scarcity present open research challenges for future intelligent transportation systems.