Details are presented of a novel approach to expedite accurate simulation of advanced power plants using CAPE-OPEN compliant Neural Network (NN) components that are derived from co-simulations involving the use of computational fluid dynamics (CFD) and process modeling software. The developed NN model is based on the well-known logistic activation function and a hybrid of simulated annealing and conjugate gradient training algorithms. In the approach discussed in this paper, the model input and output conditions required by the NN formulation are generated in precursor simulations using commercially available Aspen PlusŪ process modeling software, FLUENTŪ CFD software, and the APECS (Advanced Process Engineering Co-Simulator) interfacing software that was developed through a U.S. Department of Energy (DOE)-funded project. To ensure flexible reuse, integrity, and scalability of the simulation data, co-simulation inputs and outputs that are required for the NN training are stored in the MYSQL open-source relational database. The principal advantage of deriving a NN model, or indeed any reduced-order model, from a device-scale CFD model, as opposed to regular use of a CFD simulation model, is that depending on the range of applicability of the trained NN model, the analyst can subsequently avoid running the CFD model to reduce the turnaround time of coupled simulations.
A key feature of the current implementation is the capability to switch between the reduced-order NN model and CFD simulations to augment the solution database used for NN model re-calibration. Switching between these two modes during plant co-simulation is based on a user-specified solution strategy. The developed model is applied to two cases: a FutureGen power and hydrogen production plant, and an industrial natural gas-fired combined-cycle power plant. The predictions of the NN model in concert with Aspen Plus are found are to be in good agreement with the results of the precursor co-simulations that employ the full-fledged CFD solutions. This investigation demonstrates that co-simulations utilizing a well-trained NN component model can be conducted at a fraction of the cost of employing CFD device-scale models, without sacrificing accuracy in an engineering context. The results support the emerging consensus in the industry that co-simulation based on model order reduction is a cost effective option to achieve sustainable improvement and operation of power plants via computer simulation.