APPLICATION OF LINEAR MULTIPLE MODEL PREDICTIVE CONTROL (MMPC) FRAMEWORK TOWARDS DYNAMIC MAXIMIZATION OF OXYGEN YIELD IN AN ELEVATED-PRESSURE AIR SEPARATION UNIT

摘要

In a typical air separation unit (ASU) utilizing either a simple gaseous oxygen (GOX) cycle or apumped liquid oxygen (PLOX) cycle, the flowrate of liquid nitrogen (LN2) stream connectinghigh-pressure and low-pressure ASU columns plays an important role in the total oxygen yield.It has been observed that this yield reaches a maximum at a certain optimal flowrate of LN2stream. At nominal full-load operation, the flowrate of LN2 stream is maintained near thisoptimum value, whereas at part-load conditions this flowrate is typically modified in proportionwith the load-change (oxygen demand) through a ratio/feed-forward controller. Due tononlinearity in the entire ASU process, the ratio-modified LN2 flowrate does not guarantee anoptimal oxygen yield at part-load conditions. This is further exacerbated when processdisturbances in form of "cold-box" heat-leaks enter the system. To address this problem ofdynamically maximizing the oxygen yield while the ASU undergoes a load-change and/or aprocess disturbance, a multiple model predictive control (MMPC) algorithm is proposed. Thisapproach has been used in previous studies to handle large ramp-rates of oxygen demand posedby the gasifier in an IGCC plant. In this study, the proposed algorithm uses linear step-response"blackbox" models surrounding the operating points corresponding to maximum oxygen yieldpoints at different loads. It has been shown that at any operating point of the ASU, the MMPCalgorithm, through model-weight calculation based on plant measurements, naturally andcontinuously selects the dominant model(s) corresponding to the current plant state, whilemaking control-move decisions that approach the maximum oxygen yield point. Thisdynamically facilitates less energy consumption in form of compressed feed-air compared to asimple ratio control during load-swings. In addition, since a linear optimization problem issolved at each time step, the approach involves much less computational cost compared to a firstprinciplebased nonlinear MPC.