To address trip frequency and reduce it by 50%, Merkle was tasked with building a predictive model to identify the contributors of a trip event, then alert operators of a potential occurrence so they could take mitigating action.
Merkle’s first challenge was creating an environment that could ingest and process significant volumes of data to then feed into a live monitoring model. This data came from over 400 sensors monitoring different stages of the refinement process, many taking measurements every minute or more. Working with Spirit Energy, it began feature engineering to translate the data points, resulting in the creation of 22,658 features every minute for over three years’ worth of data, equating to 30 billion data points.
The second challenge was to create a target variable for the modelling. This was difficult due to issues around being able to identify an event trip and a non-event since not all trips had been recorded historically. Using a variety of visualisations, Merkle worked with subject matter experts at Spirit Energy to create a target variable that correctly identified an event trip and a non-event.
Finally, Merkle applied machine learning to find the signals from the noise in 22,000 features and reduced the explanatory variables down to 16. Using a random forest classification approach, Merkle then trained the model to give a propensity every minute of the likelihood of a trip happening. It then put the model into production with data scored and visualised within two minutes of it being collected.
The model was a huge success – Spirit Energy was able to identify eight out of 10 trip events at the plant, giving operators a minimum of 13 minutes advanced warning to take appropriate action. As well as having financial benefits, it is also helping ensure safety for operators.