What question did this study set out to answer?

The aim is to improve crowd forecasting accuracy using machine learning and deep learning models at Scheveningen Beach.

April 1, 2026Transportation Research Record Journal of the Transportation Research Board0 citationsOpen Access

Evaluation of Machine Learning and Deep Learning Models for Multi-Horizon Crowd Forecasting at Scheveningen Beach, The Netherlands

THTheivaprakasham HariDelft University of Technology WDWinnie DaamenDelft University of Technology SHSascha Hoogendoorn-LanserMaterials innovation institute

Key Points

The aim is to improve crowd forecasting accuracy using machine learning and deep learning models at Scheveningen Beach.
Integrated historical crowd data, weather, and holidays to train various models.
Evaluated machine learning models including CatBoost, LightGBM, LSTM, and Temporal Fusion Transformer.
Assessed model performance using multiple metrics such as mean absolute error and coefficient of determination.
CatBoost excelled in short-term forecasting, while LightGBM and LSTM performed best in mid and long-term forecasts, respectively.
Forecast accuracy tended to decrease as the time horizon increased across most locations.
Individual location models outperformed the unified models, indicating higher precision at the cost of computational effort.

Abstract

Efficient crowd management is crucial for municipalities to ensure public safety and enhance visitor experience, particularly in tourist-centric areas, such as Scheveningen Beach. Scheveningen Beach faces challenges because of the limited precision of visitor count data and the lack of accurate forecasts. Currently, crowd safety managers use their professional experience to forecast based on factors such as weather, events, and holidays, leading to inaccuracies, highlighting the need for accessible data and advanced analytics to enhance crowd management strategies. This study evaluates machine learning and deep learning models for multi-horizon hourly pedestrian crowd count forecasting, addressing the limitations of current manual prediction methods. Historical crowd data, weather, and holidays were integrated to train eXtreme gradient boosting, categorical boosting (CatBoost), light gradient boosting machine (LightGBM), long short-term memory (LSTM), and Temporal Fusion Transformer models for short-term (1-day), mid-term (7-day), and long-term (30-day) horizons. Models were developed for individual locations and as a unified multilocation approach. Performance was assessed using the coefficient of determination, root mean square error, normalized root mean square error, symmetric mean absolute percentage error, mean absolute error, and normalized mean absolute error metrics. The results showed that CatBoost was best for short-term forecasts, CatBoost and LightGBM for mid-term forecasts, and LSTM and LightGBM for long-term forecasts. Forecast performance decreases over longer time horizons in many locations, suggesting different applications: short-term forecasts for immediate operational decisions and long-term predictions for general trend analysis and strategic planning. Individual location models generally outperformed the unified approach, but at a higher computational cost. This study reveals significant spatial and temporal variability in crowd dynamics, which is crucial for optimizing resource allocation and enhancing preparedness in crowd management at Scheveningen Beach and similar tourist destinations.

Perguntar à IA

Bookmark

View Full Paper