Operator decision aid design via multi-dimensional time-series event prediction: A hydrocracking unit application

Refinery upsets lead to lower quality product, process shutdowns or equipment failure. Human operators monitor refinery processes and intervene if needed to prevent such upsets or reduce their effects. The equipment required for these processes are large and complex. There are usually many sensors to pay attention to, making the monitoring job difficult. In this work, we propose a data-driven methodology to develop operator decision aids that aim to decrease the monitoring burden. The decision aid raises soft-alarms to notify the operator that an undesirable event will happen within a certain time-frame. The event definition can be process specific (e.g. more than 5 degrees temperature rise within 15 minutes) or generic (e.g. the sensor measurement will be anomalous within 15 minutes). The idea is to predict these events using machine-learning models trained from historical data. The sensory information for a given equipment unit is treated as a multi-dimensional time-series. The events are predicted either by time-series regression and applying the event logic on the forecasts or by classification of whether there will be an event or not in a given time-frame. These events are usually rare which makes them hard to predict. This results in the trade-off between capturing all the events (ratio of correctly predicted events over total number of events) versus capturing the true events (correctly predicted events to all the positive predictions) or the trade-off be-tween recall/sensitivity and precision. If the decision aid captures all the events but mispredicts many more, the trust of the operator will be eroded. Thus, we concentrate on getting good event prediction precision values while still keeping reasonable recalls. We evaluate our approach with multiple machine learning models on the four beds of a real refinery hydrocracker unit. Our first evaluation is applying the event definitions directly. Our results show that there is at least one method that can achieve around 0.5 or higher F0.1 for each bed, implying that building a decision aid is viable. However, there isn't a clear winner among the machine-learning methods and performance changes between different beds. Predicting events for beds, instead of sensors, did not yield significant gains. Tuning the precision-recall trade-off to the end of improving the decision aid performance, we suggest an adaptive thresholding method using the running standard deviation of the prediction errors. The evaluations show that the adaptive approach can increase the performance significantly. For the last evaluation, we directly predict the change of temperature values within the target time horizon in our regression methodology. We show that there are performance gains for a subset of our data utilizing this perspective of time series prediction.