JoWUA

Abstract

The process of energy reporting in the chemical industry has evolved from a typical bookkeeping activity into a more advanced, regulatory-oriented electronic process that demands complete auditability and legal admissibility. Data-only-based or machine learning (ML)-based models typically cannot meet such stringent demands because of their very nature and vulnerability to model drift. In this work, the problem of energy reporting is treated as a problem of systems engineering, and a novel hybrid approach is developed that distinguishes between deterministic rule-based models and ML-based support tools. The approach is based on using a rule-based model to implement physical and regulatory laws, and an ML model to detect anomalies and reconcile the data. Evaluations were conducted through a multisite implementation experiment with Python and PostgreSQL. Statistically, it is evident that while a pure machine learning algorithm generated only 62% compliance gain, the hybrid system recorded a 94.8% compliance gain. Further sensitivity analysis showed that while pure machine learning would become inaccurate at 0% under cases of failure, the hybrid system retained its accuracy at 100%. Additionally, the use of the hybrid system reduced audit readiness efforts since automatic versions of evidentiary strands were provided for every metric produced. From this study, it becomes clear that having machine learning subordinate to a deterministic rule-based gatekeeper creates the ideal combination. Therefore, chemical plants are able to incorporate machine learning systems without affecting their compliance efforts.