Home // IoTAI 2025, The Second International Conference on IoT-AI // View article

Enhancing Bike-sharing Demand Forecasting: Anomaly Detection and Feature Selection in LSTM Networks

Authors:
Pedro Nunes
José Santos

Keywords: LSTM; Bike-Sharing; Feature permutation; Anomaly detection; Interpretability.

Abstract:
Accurate forecasting of casual bike-sharing demand is crucial for optimizing operations and resource allocation. This study employs a Long Short-Term Memory (LSTM) network to predict hourly bike rentals, incorporating temporal, meteorological, and categorical features. To enhance the model, we integrate an anomaly detection step using the Local Outlier Factor (LOF) method, treating its output as an additional feature. The initial LSTM model achieved a Root Mean Squared Error (RMSE) of 34.26. Incorporating anomaly detection based on weather-related data, such as temperature and humidity, and subsequently removing those features, led to an improved RMSE of 30.86. Feature permutation analysis was then used to assess variable importance. The most critical predictors were whether the day was a working day and which working day it was, highlighting clear behavioral patterns in casual bike-sharing demand. By combining anomaly detection with feature selection, we enhance the interpretability of LSTM-based forecasting models, which are often considered black boxes. Removing redundant features simplifies the model while potentially improving accuracy, making it more transparent and efficient. These findings provide valuable insights for bike-sharing system operators, enabling data-driven decision-making for demand management and operational planning.

Pages: 1 to 6

Copyright: Copyright (c) IARIA, 2025

Publication date: July 6, 2025

Published in: conference

ISBN: 978-1-68558-286-9

Location: Venice, Italy

Dates: from July 6, 2025 to July 10, 2025