Enterprise AI for Reducing Large Refinery’s Production Waste

Challenge

Automotive diesel oil (ADO) produced by the refinery is used by light duty vehicles to transport goods, requiring tight specification limits on product quality. To produce ADO, the refinery blends gas oils from various processing units. Each of the blend stocks contributes to the properties of the finished ADO and different combinations can result in different blend properties.

Variance in crude oil combinations and refining process makes it challenging to produce ADO that meets regulatory T90 specifications. As higher T90 is costly for the refinery, process engineers spend up to two days per month to produce reports that monitor and track giveaway. The reports, however, are prone to human error. Consequently, over the last three years, T90 of ADO produced at the refinery consistently exceeded regulatory T90 by 9 to 10℃.

Approach

Within 16 weeks, BHC3 experts, in collaboration with project managers and experts from the petrochemicals company, developed a production-ready application to automatically produce rigorous historical data analysis and predict ADO T90. The BHC3 team began by ingesting 36 months of data from three disparate sources to create a unified data image. The data image included data from process data historians, laboratory information systems, and historical T90 giveaway reports.

The team used the data image to configure analytics and time series expressions to perform the rigorous blending calculations. These calculations were used as inputs to machine learning models that predict ADO T90. After determining the optimal predictive model, the BHC3 team configured an application UI for operators to access AI-enabled insights and automate recurring monthly reports. With BHC3 Process Optimization, process engineers at the refinery can predict ADO T90 up to seven days in advance, identify the top contributing features associated with the T90 giveaway, and implement parameter adjustments to minimize the T90 giveaway.

Project Objectives

• Create a unified, federated data model integrating data from three disparate data sources
• Deliver a user-friendly application that automates monthly reporting and provides greater flexibility for analyzing different time windows and delivers recommendations
• Deliver machine learning models to predict upcoming T90 of automotive diesel oil blend
• Identify contributing features to adjust process conditions to reduce T90 giveaway

About the Petrochemicals Company

• $10 billion in annual revenue
• Operates in 15+ locations globally
• 5000+ employees

Project Highlights

• 16 weeks from project kickoff to pre-production application
• Ingested 36 months of data from three disparate data sources including process data historians, laboratory information management system,
  and archived T90 giveaway reports
• Configured compound metrics for rigorous blending calculations
• Configured and tested 100+ features in 10 unique machine learning models
• Configured data connectors to data sources to automate monthly giveaway reports