(152aw) Process Optimization of Ann-Based Vpsa Process for Ethane/Ethylene Separation

Pressure swing adsorption (PSA) one of separation process unit gives rise to replacement of cryogenic distillation process demanding high energy consumption. The development of separation of light hydrocarbons ethane/ethylene is achieved using vacuum pressure swing adsorption (VPSA) on metal-organic framework (MOF) sorbent for obtaining selective paraffin removal. The objective of the research is to minimize energy cost that is required for a vacuum pump of blowdown and purge step and a compressor of rinse step on vacuum pressure swing adsorption (VPSA) and to maximize ethylene recovery which is linked with minimizing ethylene productivity. To find optimal operating conditions for complex mathematical model on PSA, decision variables plays a significant role with ethylene purity holds over 99.6% with cyclic steady state (CSS). However, it requires a set of partial differential and algebraic equations (PDAES) which demands high computational time and complexity of the PSA process. A direct optimization which is a simulation model of VPSA separation process between software gPROMS and MATLAB which performs an optimization to find optimal point for ethylene productivity and energy efficiency has been performed using Particle Swarm Optimization (PSO) with 10 particles and 30 iteration that are equal to 300 points which took almost 166.53 hours to reach cyclic steady state (CSS) of PSA process. To reduce its computational time and the problem complexity, a surrogate model facilitates more rapid optimization for different multi-objectives and their pareto front as compared to the direct optimization.

The proposed method makes use of artificial neural network (ANN)-based alternative simulation model that facilitates less computational time with Latin hypercube sampling (LHS). The LHS algorithm allows to obtain a good mapping of the search area for the low correlation of decision variables. Even if the sampling took a similar computation time with the direct optimization, the surrogate model takes an advantage of resulting in different optimum points with various objective functions. To make a rigid surrogate model accuracy, hyperbolic-tangent of activation function, 4 hidden layers, 7 neurons and Levenberg-Marquart of back propagation were used. PSO with time varying acceleration coefficient (TVAC)-based technique were used to have optimum point on pareto front. Surrogate model shows a reliable performance when model accuracy is around 1e-05 that was checked k-fold validation with 5 folds instead of setting test data positioned in training set normally verified in ANN model. Other 50 samples rather than LHS method were proposed using Euclidean method around optimum point of PSO optimization. The optimum value for ethylene recovery has been changed from 68.1% to 80.148% which is base-conditioned point with respect to PSA column length over diameter ratio throughout optimization. It is shown that VPSA alternative process is economically competitive based on optimized energy consumption (MJ/kg(ethylene)) for the VPSA steps and ethylene liquification after VPSA process, as compared to 0.77 MJ/kg from the distillation process.