An evolutionary programming technique for evaluating the effect of ambient conditions on the power output of open cycle gas turbine plants - A case study of Afam GT13E2 gas turbine

Abstract

In this research a novel Evolutionary Programming (EP) approach combined with a Continual Learning Neural Network (EP-CLNN) technique is proposed for the evaluation of Open Cycle Gas Turbines (OCGTs) net power output considering ambient temperature conditions. Studies were performed comparatively considering the proposed EP-CLNN approach, state-of-the-art linear regressors of several Polynomial (Poly) orders (Poly-1, Poly-2, Poly-3, and Poly-4) and three relatively small-to-big datasets. These approaches were applied to two open source datasets reported in the literature and a case study small dataset acquired from the Afam gas power station plant (GT132E2). The results showed that the reported Root-Mean Squared Error (RMSE) and mean net power estimates for the EP-CLNN are indeed promising considering the constraint of sequential continual learning prediction requirement, reduced variable (single input variable), simpler function set, and relatively small datasets. For the case of the big open source dataset (Dataset-1), the proposed EP-CLNN approach gave the best solution with least RMSE of 4.9926 MW over the linear regressors - Poly-1, Poly-2, Poly-3, and Poly-4 with RMSE estimates of 5.4609 MW, 5.2305 MW, 5.0986 MW, and 5.0980 MW respectively. For the case of the small dataset (Dataset-2), the estimated mean net powers were 32,974.0000 MW, 59140.0000 MW, and 96.2143 MW for Poly-1, Poly-2, and the EP techniques respectively; when compared to the base (reference) net mean power of 97.52 MW, the EP technique is the better one. Furthermore, for dataset 3, the EP technique gave an RMSE of 0.3381 MW while the combined solution of EP-CLNN gave a Mean Absolute Percentage Error (MAPE) consensus estimate of 0.1377 from a synthesized dataset of 20 evolved symbolic expressions. Thus, the EP-CLNN represents a promising approach to real-time gas turbine power output modeling with better accuracy and unique consensus expression capability.