Successful Machine Learning Strategies in an Environment of Intermittent Data Availability

This report introduces an online, fully adaptive HybridANN approach for the estimation of product properties across multiple crude distillation processes. An extension of eTS+ is considered, utilizing recursive density estimation (RDE), eClustering+ and fuzzily weighted recursive leastsquares (fwRLS) for one-step prediction of fuel quality in a data-sparse, sensor-heavy environment.[49] eTS+ has been improved with mutual information based input variable selection, online monitoring of regressor utility, and online normalization to produce a system evolving in both features considered and number of rules for autonomous application to any refinery. The solution has only optional requirement for user-defined parameters and expert knowledge, and produces a transparent model for product estimation.