Signal Decomposition - Driven Deep Learning Framework for Passenger Demand Forecasting: Evidence from Dublin Luas

Abstract

The Dublin Luas is an urban light rail system that exhibits complex and dynamic passenger demand, which poses significant challenges for planning and scheduling decisions. Effective management of the system requires reliable passenger demand forecasts across short-, medium-, and long-term horizons to support operational control, scheduling, and capacity investment decisions. This study proposes a hybrid Artificial Intelligence (AI) framework that integrates signal processing and deep learning for multi-horizon passenger demand forecasting. The framework first applies Variational Mode Decomposition (VMD) to the weekly passenger demand series from the Dublin Luas Red and Green lines. This decomposes the data into distinct temporal components. These components are then modelled with Bidirectional Gated Recurrent Units (BiGRU) to capture their frequency-specific temporal dependencies. The BiGRU hyperparameters are optimized using the Tree-structured Parzen Estimator (TPE) algorithm. Results demonstrate that the proposed hybrid VMD+BiGRU framework outperforms the baseline models. On the Red Line, it achieved an RMSE of 7.88, an MAE of 5.85, and an R2 of 0.976. On the Green Line, it achieved an RMSE of 9.65, an MAE of 8.47, and an R2 of 0.958. Wilcoxon signed-rank tests further confirmed the superiority of the proposed framework. To further evaluate model robustness and temporal transferability, a multi-horizon forecasting analysis was conducted for short-term (1-week), medium-term (4-week), and long-term (12-week) intervals. The VMD+BiGRU model maintains low error growth and high predictive accuracy across all horizons compared with competing models. The proposed framework provides reliable forecasts that support resource allocation, timetable adjustment, and operational management of the Dublin Luas system.