Neural nets for thermodynamic properties

Abstract

Thermal power plants are controlled and monitored in real time using state-of-the-art distributed control systems (DCS). Both at real-time control and online monitoring of these plants, thermodynamic properties of water/steam along with their partial derivatives are required in computation and optimization of process. Particularly during data reconciliation and process optimization, the equipment models appear as non-linear constraints which require Jacobians/Hessians of thermodynamic properties. Thermodynamic properties of steam/water are complex functions (Gibbs and Helmholtz functions), and they are repeatedly called during non-linear optimization. The Jacobians/Hessians of these properties are numerically approximated involving further repeated calls to the functions. Legacy approaches involve approximating these functions using higher-order polynomial algorithms or look-up tables. These methods have the limitation of high computational time or need large memory. However, using artificial neural networks (ANNs), these limitations can be overcome, and in this paper, it is demonstrated that ANNs with back-propagation neural networks (BPN) are an effective means to improve the computational performances. The computational time is reduced by a factor of nearly four times.