Data Reduction and Reconstruction of Wind Turbine Wake Employing Data Driven Approaches

Data driven approaches are utilized for optimal sensor placement as well as for velocity prediction of wind turbine wakes. In this work, several methods are investigated for suitability in the clustering analysis and for predicting the time history of the flow field. The studies start by applying a proper orthogonal decomposition (POD) technique to extract the dynamics of the flow. This is followed by evaluations of different hyperparameters of the clustering and machine learning algorithms as well as their impacts on the prediction accuracy. Two test cases are considered: (1) the wake of a cylinder and (2) the wake of a rotating wind turbine rotor exposed to complex flow conditions. The training and test data for both cases are obtained from high fidelity CFD approaches. The studies reveal that the combination of a classification-based machine learning algorithm for optimal sensor placement and Bi-LSTM is sufficient for predicting periodic signals, but a more advanced technique is required for the highly complex data of the turbine near wake. This is done by exploiting the dynamics of the wake from the set of POD modes for flow field reconstruction. A satisfactory accuracy is achieved for an appropriately chosen prediction horizon of the Bi-LSTM networks. The obtained results show that data-driven approaches for wind turbine wake prediction can offer an alternative to conventional prediction approaches.

Automated summary

Data driven approaches are used for optimal sensor placement and velocity prediction in wind turbine wakes. Various methods are investigated for clustering analysis and predicting the flow field's time history. The studies show that a combination of classification-based machine learning algorithm and Bi-LSTM can predict periodic signals accurately, while a more advanced technique is needed for complex turbine near wake data.