Dynamic modeling of batch polymerization reactors via the hybrid neural-network rate-function approach

A simulated verification and validation of the hybrid neural-network rate-function (HNNRF) approach to modeling batch reactor systems is provided. In chemical reactor processes, some measurements may not be easily obtainable, and the designed neural-network rate-function (NNRF) model in our previous work did not propose a method to include the state variables for the suggested dynamic model. To overcome this difficulty, the approximated mechanistic equations characterizing these immeasurable state variables could be incorporated into the NNRF model to form the hybrid neural-network rate-function model. The sequential pseudo-uniform design (SPUD) is used to locate the sufficient but limited experiments to provide the HNNRF model with rich information. In this research, the HNNRF modeling capability over a large operating region was evaluated employing a simulated polymerization reactor system. In addition to the comparative benefit of short time expenditure for building the model, the performance of the identified HNNRF model is quite acceptable in the face of noisy measurements and the identified model could be applied to determine the optimal recipe or the operating conditions of the reactor systems.