(362b) Multi-objective Optimization of Process Efficiency and CO2 Emission for On-Site Hydrogen Production Process Using Hybrid Data-Driven Model.
AIChE Annual Meeting
Tuesday, November 15, 2022 - 3:30pm to 5:00pm
For robust multi-objective optimization of the SMR process, a hybrid deep neural network (DNN) model was developed. The hybrid DNN model consists of an operation DNN trained with plant data and a simulation DNN trained with simulation data. The operation DNN was evaluated using R2 and NRMSE from the predicted and actual values and showed high prediction accuracy, 0.94 and 3.89 on average, respectively. Further, the simulation DNN was able to soft-sense the variables that could not be acquired from the operation DNN since it was unable to collect data from an actual process. In addition, the hybrid DNN model improved the prediction reliability through cross-validation of output variables between operation DNN and simulation DNN.
Multi-objective optimization was performed to find Pareto optimal solutions using the multi-objective particle swarm optimization (MOPSO) algorithm. The Pareto front showed a process efficiency between 77.5 and 87.0% and CO2 emission between 577.9 and 597.6 t/y. Further, several representative solutions have analyzed the relationship between the optimal solutions in the Pareto front. Decision-makers can select appropriate solutions from these representative solutions to enable flexible operations with respect to specific operational requirements or safety guidelines.