(430d) Advanced Modeling and Optimization for Future Generation Energy Systems

Optimization models are essential tools for decision-making. They are encountered in all facets of process engineering, from model and process development, to process synthesis and design, and for process operations, control, scheduling and planning. Unfortunately, engineering practice views most optimization strategies as search techniques that are wrapped around a simulation model. This glorified case study approach can be very expensive to execute, and it is often not clear whether excellent solutions are overlooked.

A more efficient approach is to directly solve the optimality conditions as an extension of the process model. This nonlinear programming (NLP) approach ensures that the best solutions have been found and that firm knowledge of neighboring solutions can be assessed. Moreover, sensitivity of the optimal solution to exogenous inputs, and robustness of the solution to uncertainties can be assessed with negligible cost through sensitivity analysis. Finally, an efficient and reliable optimization strategy integrates well within the work process and easily adapts to model variations and extensions. These features are addressed by leading edge process optimization modeling environments. Embedded within state-of-art optimization modeling tools, these approaches provide optimal solutions to real-world problems at far less cost (often orders of magnitude) than simulation-based approaches. In particular, a recent optimization framework has been developed by the Institute for the Design of Advanced Energy Systems (IDAES) at NETL, using Pyomo, a Python-based optimization modeling environment. The resulting platform leads to a powerful decision-making capability for large, complex energy systems.

Armed with these capabilities, this talk discusses recent advances in the integration of optimization into the engineering workplace. First, equation-oriented optimization models couple state of the art NLP solvers to parallelizable structures, and realize fast and reliable convergence behavior, in contrast to simulation-based optimization approaches. This has led to the solution of problems with potentially millions of variables and thousands of degrees of freedom. Second, enabling NLP tools handle optimization models with nonsmooth switches as well as phase changes in equilibrium systems. This capability allows the modeling of phase changes and complex phenomena in distillation columns, pipelines, complex heat exchangers and reservoir models. Finally, detailed (and expensive) procedures for device-scale (e.g., CFD) and physical property models derived from computational chemistry (DFT, molecular dynamics) resist the integration within system-wide optimization models. Instead, reduced models (RMs) are often substituted for decision-making tasks, but often with compromises on the optimality of these decisions. Recently, RM-based trust region frameworks have been developed that guarantee convergence to the optimum of the original detailed model (ODM), through the solution of RM-based trust region strategies with recourse to ODM evaluations. All of these features will be illustrated by challenging optimization case studies presented in the talk.