A Three-Layer LSTM Approach for Proactive Train Schedule Variability Alerts for Individuals with Reduced Mobility

Individuals with reduced mobility rely on accurate information to navigate transit systems effectively. Proactive notifications about Train Schedule Variability (TSV) can significantly enhance their travel experience by providing timely alerts on potential delays. This paper introduces a three-layer...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IEEE/IFIP Network Operations and Management Symposium s. 1 - 10
Hlavní autoři: Dardour, Mayssa, Mosbah, Mohamed, Ahmed, Toufik
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 12.05.2025
Témata:
Accuracy
Delays
Edge computing
Fault detection
Hierarchical Clustering Approach (HCA)
Indexes
Long short term memory
LSTM
Predictive models
Reduced mobility
Schedules
Through-silicon vias
Time series analysis
Train Schedule Variability (TSV)
Variability Assignment Score (VAS)
Variability Index (VI)
Wheelchairs
ISSN:2374-9709
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Individuals with reduced mobility rely on accurate information to navigate transit systems effectively. Proactive notifications about Train Schedule Variability (TSV) can significantly enhance their travel experience by providing timely alerts on potential delays. This paper introduces a three-layer Long Short-Term Memory (LSTM) model to predict TSV and notify reduced-mobility individuals through a Human-Machine Interface (HMI) system integrated into their wheelchairs. Our model estimates a Variability Index (VI) over 7, 15, and 30-day windows within a hierarchical architecture spanning edge, fog, and cloud layers. The edge layer delivers advance notifications, alerting users up to 7 days prior to anticipated delays. A Hierarchical Clustering Approach (HCA) is applied for robust data preparation, addressing missing values and identifying weekly patterns in a one-year dataset of train schedules from Frankfurt Station. This dataset, inclusive of seasonal and weather variability, undergoes additional processing with median filters and time-sliding windows to create data matrices. We evaluate the proposed three-layer LSTM model against multiple LSTM configurations and traditional time series forecasting models, including AutoRegressive Integrated Moving Average (ARIMA), Kalman Filtering, and an ARIMA-Kalman hybrid. Model performance is assessed using Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and Mean Absolute Error (MAE). Additionally, results demonstrate that the Variability Assignment Score (VAS) surpasses Mean Squared Error (MSE) in fault detection accuracy. Receiver Operating Characteristic (ROC) curve analysis, using Area Under the Curve (AUC) metrics, validates the superior performance of our model. Specifically, the VAS-based ROC curve for the three-layer LSTM model outperforms others, underscoring VAS as a more effective measure for variability detection. Further validation via the Diebold-Mariano (DM) test confirms the suitability of the three-layer LSTM model for this application.
ISSN:2374-9709
DOI:10.1109/NOMS57970.2025.11073713