(753c) A Hybrid Modeling Architecture
AIChE Annual Meeting
2019
2019 AIChE Annual Meeting
Computing and Systems Technology Division
Data-Driven and Hybrid Modeling for Decision Making
Friday, November 15, 2019 - 8:38am to 8:57am
The general problem that a hybrid model needs to solve, is the prediction of multiple output variables such that the predictions agree well with measurement data while not violating any first principles. There are currently four main architectures for the creation of hybrid models. The first one is a parallel architecture in which a whiteâbox (WB) model and a blackâbox (data-driven) model are used in parallel to make predictions and then some sort of prediction averaging is carried out to predict the desired output variables. However, this approach does not guarantee no violation of first principles. The parallel architecture can also be used to predict some output variables using a WB model and to predict the remaining output variables using a black-box (BB) model. The second one is a WB/BB serial architecture in which a WB model and a BB model are used in series to make predictions. The third one is a BB/WB serial architecture in which a BB model and a WB model are used in series to make predictions. Finally, the fourth one is a WB-embedded-in-BB architecture, in which known physical and chemical laws are embedded in a BB model as equality and inequality constraints.
In this paper, we propose and implement a hybrid model architecture in which BB models are used to predict unknown/uncertain parameters and inputs of a WB model. The WB model then predicts desired output variables based on the BB-model predictions and values of known/certain inputs. The application and performance of this hybrid model architecture are shown using numerically-simulated chemical process examples.