A Comparative Evaluation of Time Series Forecasting Models for Pedestrian Footfall Prediction in Dublin City Centre

Abstract

This study conducts an in-depth comparative evaluation of time series forecasting methodologies applied to pedestrian footfall data from Dublin City Centre. Four distinct models are examined: Holt-Winters Exponential Smoothing, Seasonal Autoregressive Integrated Moving Average (SARIMA), Facebook’s Prophet and Long Short-Term Memory (LSTM) neural networks. Using Data Collected via PYRO-BOX sensors across the 15 streets from 2022 to 2024, this study conducts a comparative evaluation of time series forecasting models to address limitations in previous studies, which often focus on single methods or fail to capture the non-linear dynamics of urban pedestrian movement. The study implements an 80/20 train-test split and evaluates model performance using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-squared metrics. The findings indicate that the LSTM model performs considerably better than conventional methods, achieving the lowest prediction errors (MAE: 26,069.72; MAPE: 8.82%) and an R² of 0.7778. The Prophet model emerges as a strong alternative, balancing predictive accuracy with interpretability with a MAPE value of (12.78%). Traditional statistical models show limitations in capturing the non-linear dynamics and irregular patterns common in urban pedestrian movement. These findings provide actionable insights for model selection in real-world forecasting applications and contribute to the growing body of knowledge in urban analytics and smart city infrastructure development.