A novel dynamic wind farm wake model based on deep learning

Abstract

A deep learning based reduced order modelling method for general unsteady fluid systems is proposed, which is then applied to develop a novel dynamic wind farm wake model. The proposed method employs the proper orthogonal decomposition technique for reducing the flow field dimension and the long short-term memory network for predicting the reduced representation of the flow field at a future time step. The method is specifically designed to tackle distributed fluid systems (such as wind farm wakes) and to be control-oriented. For wind farm wake modelling, a set of large eddy simulations are first carried out to generate a series of flow field data for wind turbines operating in different conditions. Then the proposed method is employed to develop the data-based wake model. The results show that this novel dynamic wind farm wake model can predict the main features of unsteady wind turbine wakes similarly as high-fidelity wake models while running as fast as the low-fidelity static wake models and that the model’s overall prediction error is just 4.8% with respect to the freestream wind speed. As an illustrative example, the developed model can predict the unsteady turbine wakes of a 9-turbine test wind farm within several seconds based on a standard desktop while it requires tens of thousands of CPU hours on a high-performance computing cluster if a high-fidelity model is used. Thus the developed model can be used for fast yet accurate simulation of wind farms as well as for their predictions and control designs.