Process Modeling using Structured Neural Networks

A modeling approach is developed which combines a partial first principles model, incorporating the available prior knowledge about the process being modeled, with a neural network which serves as a non-parametric estimator of unmeasured process parameters that are difficult to model. This hybrid model is superior to standard "black-box" neural network models in that it interpolates and extrapolates much more accurately, is easier to analyze and interpret, and requires significantly fewer training examples. The hybrid network model, when used to model a fedbatch bioreactor, gives estimates of the unobserved process parameters and can be used to make predictions. This approach can also be applied when only part of the state is measured by using a state reconstruction method for the first principles component of the hybrid model.