Hybrid Model of an Evaporative Gas Turbine Power Plant Utilizing Physical Models and Artificial Neural Networks

Traditionally, when process identification, monitoring and diagnostics are carried out for power plants and engines, physical modeling such as heat and mass balances, gas path analysis, etc. is utilized to keep track of the process. This type of modeling both requires and provides considerable knowledge of the process. However, if high accuracy of the model is required, this is achieved at the expense of computational time. By introducing statistical methods such as Artificial Neural Networks (ANNs), the accuracy of the complex model can be maintained while the calculation time is often reduced significantly reduced. The ANN method has proven to be a fast and reliable tool for process identification, but the step from the traditional physical model to a pure ANN model is perhaps too wide and, in some cases, perhaps unnecessary also. In this work, the Evaporative Gas Turbine (EvGT) plant was modeled using both physical relationships and ANNs, to end up with a hybrid model. The type of architecture used for the ANNs was the feed-forward, multi-layer neural network. The main objective of this study was to evaluate the viability, the benefits and the drawbacks of this hybrid model compared to the traditional approach. The results of the case study have clearly shown that the hybrid model is preferable. Both the traditional and the hybrid models have been verified using measured data from an existing pilot plant. The case study also shows the simplicity of integrating an ANN into conventional heat and mass balance software, already implemented in many control systems for power plants. The access to a reliable and faster hybrid model will ultimately give more reliable operation, and ultimately the lifetime profitability of the plant will be increased. It is also worth mentioning that for diagnostic purposes, where advanced modeling is important, the hybrid model with calculation time well below one second could be used to advantage in model predictive control (MPC).