Steam Turbine Exhaust Optimization Based on Gaussian Covariance Networks Using Transient CFD Simulations

Renewable energies are increasingly contributing to the overall volume of the electricity grid and demand besides high efficiency, greater flexibility of the conventional fossil power plants. To optimize these objectives, extensive CFD calculations are required in most cases. For example, transient CFD calculations are only rarely combined with an optimizer because of their high demand on computational resources and time. Surrogate models, which are mathematical methods to learn and approximate the relationship between input and output parameters, are a common way to solve these problems. Once they are trained, they can perform the evaluations within seconds and replace the expensive simulation. Of course, real calculations are still needed to generate the training data. Therefore, it is useful to apply efficient and sequentially extensible design plans. This paper presents a new surrogate model method, based on a deep neural network learning the non-stationary hyperparameters of combined Gaussian process covariance matrices. It is used to approximate the complex and time consuming transient CFD simulation of a combined high-intermediate pressure steam turbine double shell outer casing. To minimize the exergy loss, the exhaust geometry is optimized in a single and multi-objective optimization on the surrogate models. The multi-objective optimization also includes the uniform velocity distribution of the steam in different areas of the casing, to predict the thermal loading of the steam turbine inner casing and to avoid an imbalanced thermal loading. A sequential sampling approach combined with a sensitivity analysis is used to find the minimum number of samples needed to train the surrogate models in order to gain sufficient prediction quality. Additionally, the paper describes the initial geometry, its numerical setup and the required control mechanisms to avoid noisy designs, which might complicate the surrogate model training. There is also a comparison of the initial and chosen optimal designs.