Autonomous Refinery Intelligence Architecture

Abstract

Abstract Refinery decisions often require evidence that is distributed across plant historians, maintenance systems, work permits, incident records, management-of-change registers, audit actions, shift logs, engineering documents, and workforce-compliance records. Conventional search or dashboard solutions generally address one source type at a time and leave engineers to assemble the complete operational story manually. This paper presents the Autonomous Refinery Intelligence Architecture (ARIA), an on-premise, evidence-grounded architecture developed in-house at a public-sector refinery to unify structured plant records and unstructured operational knowledge while maintaining data sovereignty. The architecture is designed to connect 15 enterprise systems through governed tools and a Model Context Protocol integration layer. A refinery-wide source inventory contained 423,614 structured records and 6,898 shift-log records across 12 operational domains. The detailed case study evaluates eight operational source classes using controlled subsets from crude and vacuum distillation and hydroprocessing units. The method combines canonical equipment identity, deterministic source-quality checks, structured and chronological candidate generation, retrieval-augmented language-model synthesis, a refinery-specific qualification gate, and human review. Three domain patterns are supported: safety investigation, reliability investigation, and contractor-workforce compliance. The language model acts as a controlled reasoning and summarization layer; it is not permitted to authorize plant or workforce decisions. The evaluated crude-distillation pipeline assembled 52,664 evidence rows and reduced 124 candidate packs to 15 final cards, of which only three met the strongest machine qualification. The hydroprocessing pipeline assembled 81,313 rows and reduced 132 candidate packs to 15 cards, again with three receiving the strongest qualification. The deliberate rejection and review rates show that conservative evidence gating is more important than maximizing the number of generated insights. The evaluation also exposed material source-data limitations, including missing work-permit tags, invalid notification dates, and incomplete closure fields. These limitations were carried forward as confidence constraints rather than silently repaired by the model. An initial internal workflow benchmark for a representative incident investigation indicated approximately 80% less evidence-assembly time than the prior manual approach. This is a preliminary operational benchmark, not a controlled statistical trial. A 15-20% reduction in unplanned downtime remains a projected benefit to be tested through longer-term asset performance monitoring. The principal conclusion is that agentic artificial intelligence can support refinery operations only when identity, provenance, qualification rules, human accountability, and on-premise governance are designed as first-class system components.