A deep-learning approach for 3D realization of mean wake flow of marine hydrokinetic turbine arrays

We present a novel convolutional neural network (CNN) algorithm to reconstruct turbulence statistics in the wake of marine hydrokinetic (MHK) turbine arrays installed in large meandering rivers. To train the CNN, we utilize large eddy simulation (LES) data depicting the wake flow from a single row of turbines. Once trained, the CNN is deployed to forecast the wake flow of MHK turbine arrays under different hydrodynamic conditions and for varying waterway plan-form geometry. Validation of the CNN predictions are conducted using independently performed LES. Our findings demonstrate the capacity of CNN to accurately predict the wake flow of MHK turbine arrays at significantly reduced computational cost compared to LES. Additionally, the comparison between CNN and unsteady Reynolds-averaged Navier-Stokes (URANS) simulation exhibits a notable advantage of CNN in prediction efficiency and accuracy. This research highlights the potential of CNN to establish reduced-order models for facilitating control co-design and optimization of MHK turbine arrays within natural environments. • A novel machine learning framework based on LES is proposed to predict the MHK turbine’s wake flow. • The proposed machine learning framework can accurately predict the wake flow of utility-scale tidal farms. • The proposed method has a significantly higher accuracy than URANS model. • The proposed method costs only 2.8 % of the computational time and resources required by LES.