(24b) Extracting Value from Plant Data

In addition to their normal products, refineries and petrochemical plants produce large amounts of data – as much as 100 gigabytes a day from a typical site, with the amounts continually increasing.  But how can the value of this data be maximized? The data contains a wide mixture of time stamped data types – numeric process variables values, commercial transactions, textual information, geographic equipment location records, frequency spectrums from special laboratory equipment and rotating equipment vibration measurements, and video/ audio recordings. However, the data is natively of poor quality and is not necessarily well structured.  Process instrument readings drift and noise corrupts the measurements. Even when the actual measurements are good, the statistical properties are not – i.e. process data is usually non-stationary, serially auto-correlated and cross-correlated.  After conditioning, the data is normally processed through a model to obtain some type of performance indicator, perhaps of an individual piece of equipment or of the overall plant or site – for example, unit yields or energy consumption.  The performance indicators are then used to evaluate current and historical performance against a standard and identify gaps.  However, significant additional value and margin can be gained by improving the accuracy of forecasts of future plant behavior, including potential production and supply chain alternatives, early detection of potential equipment problems, and product quality issues.  Extensive developments in the area of predictive analytics, which involves identifying patterns and relationships in plant data and using the models developed to improve decisions and increase margins, have greatly improved the potential quality and accuracy of these forecasts.  In this presentation, actual case studies will be used to illustrate the impact of these new tools on plant productivity and margins.