Smart process analytics for predictive modeling

•An automated framework is proposed for the selection of data analytics methods.•Methods are selected based on the data characteristics and domain knowledge.•High model accuracy is observed for datasets for a variety of process systems.•For a four-stage evaporator, the long-term prediction error is half that of a recurrent neural network.•In another application, machine learning methods have up to 30% lower prediction error than PLS. [Display omitted] While data analytics tools are changing how manufacturers make critical decisions and designs, the selection of the best method requires a substantial level of expertise. In practice, methods are chosen based on familiarity or on cross-validation results from a large candidate model pool which over-fits data. A Smart Process Analytics framework is presented which empowers the users to focus on goals rather than on methods and automatically transforms manufacturing data into intelligence. The method selection is based on domain knowledge, the specific data characteristics, and nested cross-validation procedures. The approach is demonstrated in case studies for experimental datasets from a variety of process systems. For a four-stage evaporator, a state-space identification method is selected that has half the long-term prediction error than a recurrent neural network. For a combined cycle power plant, machine learning methods are selected that have up to 30% lower mean-squared error than partial least squares.