Only ~10% of incidents have enough data captured in official logs to be used to make future predictions [Alex Clark – Head of Consulting, JNCTION]
Background:
The rail industry is currently looking at an incomplete picture of the past when trying to build predictive machine learning models and make business decisions.
For example, only ~10% of incidents have enough data captured in official logs to be used to make future predictions. Timestamps for starts of incidents are missing or incorrect.
Jnction faced this same challenge when building models to predict the durations of live incidents on the East Coast Mainline.
And there is also no data-driven way of declaring an incident area ‘In Order’ after the issues are resolved – we have to trust humans have remembered to log it. The return to normal service and benchmarking Control team response is also difficult. How can we measure “return to normal service” if we don’t know what normal is?
Our Solution:
JNCTION were granted research & development funding by Innovate UK to improve the rail industry’s delay prediction capabilities.
Combining rail performance expertise and access to granular historic train movements data through JNCTION Archive, Jnction’s data scientists went back to first principles and invented a new performance metric for the rail industry which can be applied to: maintenance, operations, control, delay predictions and customer information.
Jnction’s proprietary ‘Service Level Metric’ reviews the health of a particular geography (e.g. a station/stanox) over a 30-minute period. We can then track performance over an incident duration to assess how much service was affected and how quickly it recovered.
Benefits:
Jnction’s new research and development enabled us to infer missing ‘Incident End’ timestamps where logs did not have this data. As a result, our machine learning models predicting incident durations had up to 5 times more training data, leading to enhanced accuracy.
The Service Level Metric also has applications to benchmarking service recovery, as we can define ‘normal service’ and help operators compare how long it takes to recover normal service, for different incident types.
These new data also feed into business decisions for Maintenance, including being data-driven to decide the training plans for engineers to reduce ‘fix times’.
By optimising your operations and maintenance response, you can significantly reduce your Schedule 8 and delay repay costs.
