Artificial Intelligence-oriented economic non-linear model predictive control applied to a pressure swing adsorption unit: Syngas purification as a case study

A novel Artificial Intelligence-oriented Economic nonlinear model predictive control (EMPC) strategy is proposed herein. It was applied to a Pressure Swing Adsorption (PSA) unit where syngas is purified by porous aminofunctionalized titanium terephthalate MIL-125. This study addresses the open research problem of simultaneous economic optimization and advanced control in PSA units. A surrogate model based on a Deep Learning Neural Network (DNN) architecture is built to capture the process dynamics with high fidelity. The deployment of these DNN models into the EMPC control law allows obtaining solutions at a low computational cost, making it feasible to be applied in practice, even if using a particle swarm optimization technique. The robustness of the proposed EMPC controller is evaluated under a realistic plant-model mismatch in a software-in-the-loop approach. An experimentally validated phenomenological nonlinear model of the PSA unit for syngas purification was used to represent the plant and to train the DNN models. While EMPC maximizes the unit productivity in all zone/output tracking scenarios for CO2 recovery and CO2 purity, it can successfully fulfill the key operating endpoint constraint related to the H2/CO ratio, including the unmeasured perturbation scenarios.

Automated summary

The novel EMPC strategy proposed in this study combines Deep Learning Neural Network models to optimize and control PSA units in real time, achieving high efficiency and low cost applications. Through software-in-the-loop testing, the robustness of the EMPC controller against plant-model mismatch was demonstrated. EMPC successfully maximizes CO2 recovery and CO2 purity while meeting the key operating endpoint constraint of H2/CO ratio.