Kaplan Turbine Performance Prediction Using CFD: an Artificial Neural Network Approach

In this work a response surface approach was used for predicting the hill chart of a small horizontal shaft Kaplan turbine. To this aim, artificial neural networks (ANN) were chosen as a fast, reliable, and computationally inexpensive tool. The training of the ANN was based on computational fluid dynamics (CFD). The optimization of the runner-guide vane stagger correlation is a time consuming and expensive task when obtained either on experimental models or directly on the power plant. The use of CFD coupled with ANN may represent an attractive alternative. In order to obtain a more general prediction tool it was decided to characterize the turbine performance coupled with a generic draft tube. The draft tube is a key component for the performance of small head power plants. As a matter of fact, the kinetic energy recovery may represent a considerable fraction of the total head, thus strongly affecting the efficiency. It presents a very complex flow environment characterized by unsteady, large scale vortices, and the prediction of its performance is a challenging task for the CFD. Moreover, in the refurbishment and repowering of hydro power plants, the turbine may often be coupled with a pre-existing draft tube. For these reasons the draft tube was not included in the computational domain, and a simple 1D model was used to account for its influence on the turbine. The three-dimensional version of the TRAF code developed at the University of Florence, coupled with a two equation turbulence closure has been used to predict the turbine flow features and the operating characteristics. The code exploits the artificial compressibility concept to work with incompressible flows. Steady multirow viscous single-phase analysis has been applied to compute the flow field from the inlet struts to the runner exit. The CFD results were used for the training of a feed-forward artificial neural network with two hidden layers. As far as the training is concerned, a gradient based back propagation method was employed. In order to improve the generalization ability, a hybrid network made by multiple trained neural networks was used. The considered input parameters were the stagger angles of the runner and the guide vane, the total head and the draft tube recovery coefficients. The outputs of the neural network were the computed mass flow rate and the hydro power plant efficiency. The proposed procedure was applied to an existing power plant in order to optimize the runner-guide vane stagger correlation for all runner positions. Comparisons with the measurements obtained on the hydro power plant are presented and discussed. The predicted coupled positions of the guide vane and runner blades were successfully verified all over the range of the operating mass flows.