Temporal Instability Forecasting for Early Operational Risk Prediction in Railway Systems
| dc.contributor.author | Sourabh, Shreyansh | |
| dc.date.accessioned | 2026-08-20T10:27:00Z | |
| dc.date.available | 2026-08-20T10:27:00Z | |
| dc.date.issued | 2026 | |
| dc.description.abstract | Railway dispatch systems already log nearly everything that goes wrong on a network. Delays, cancellations, the cause text of each incident, all of it sits in the database. What is a bit odd is that none of this gets rolled up into a single number that tells an operator how stressed the system actually is. Most action happens only after a disruption is big enough for passengers to feel it, and there isn’t really a clean way to look a few weeks ahead. This work tries to fill that gap. It proposes the Operational Instability Index (OII), a bounded weekly score in [0, 1], built from the openly available Belgian Railway incident data between January 2019 and January 2026 (891 records spread across 318 weeks). The OII is put together from six Min-Max normalised features: weekly incident frequency, total delay minutes and cancellation counts, plus the deviation of each from an eight-week rolling baseline. The two groups are then added as a weighted sum, with the absolute components carrying weight 1.0 and the deviation components 0.5. We chose lower weights for the deviations because that series is clearly noisier. An exponential moving average is then applied; without smoothing the lag-1 autocorrelation of the series is very low and forecasting is more or less hopeless. Before fitting anything, a natural log transform is used to cut the right skew (skewness drops from 1.91 to 0.67), and both ADF and PP tests confirm stationarity, so no differencing was needed. Four classical time-series models, ARIMA, SARIMAX with Fourier harmonics, Facebook’s Prophet, and Holt-Winters exponential smoothing, were trained on 253 weeks and evaluated using a rolling one-step-ahead scheme on a 65-week held-out window. An inverse-RMSE weighted ensemble of the four reaches RMSE = 0.0319 and MAPE = 17.57% on the original OII scale. That is the best of every configuration we tried, although the margin over the individual SARIMAX and ARIMA fits is small (and not statistically significant under the Diebold–Mariano test). The forecasts are then mapped into a three-tier probabilistic early warning scheme (Critical, High, Elevated, Low) using percentile thresholds derived only from the training period, to avoid any leakage. The result is a simple, interpretable instrument that hands operators a twelve-week probabilistic outlook on operational health, instead of yet another post-mortem after a bad week. | |
| dc.identifier.uri | http://nits.ndl.gov.in/handle/123456789/101 | |
| dc.publisher | National Institute of Technology, Silchar | |
| dc.title | Temporal Instability Forecasting for Early Operational Risk Prediction in Railway Systems |
Files
Original bundle
1 - 1 of 1
No Thumbnail Available
- Name:
- Thesis_Final_Shreyansh_Sourabh-1 - Shreyansh Sourabh (1).pdf
- Size:
- 2.97 MB
- Format:
- Adobe Portable Document Format
License bundle
1 - 1 of 1
No Thumbnail Available
- Name:
- license.txt
- Size:
- 1.71 KB
- Format:
- Item-specific license agreed upon to submission
- Description: