A measure–correlate–predict model based on neural networks and frozen flow hypothesis for wind resource assessment

Abstract

In this paper, a measure–correlate–predict (MCP) model based on neural networks (NN) and frozen flow hypothesis, which is abbreviated as the MCPNN-frozen model, is proposed for wind resource assessment and tested using turbulent channel flows with three different surface roughness lengths, i.e., [Formula: see text], 0.01, and 0.1 m. The predictions from the MCPNN-frozen model are compared with the real data for different separations ( s) between the reference point and the target point. The results show that the correlation coefficients C.C. between the model predictions and real data are roughly higher than 0.5 for small separations [Formula: see text] (where δ is the boundary layer thickness), and the coefficients of determination ( R2) are approximately higher than 0.3 when [Formula: see text]. The generalization capacity of the MCPNN-frozen model is tested for different roughness lengths and different velocity components. Further analyses show that, even though C.C. and R2 decrease when increasing s, the large-scale variations of velocity fluctuations are well captured by the MCPNN-frozen model especially for the one trained using the data filtered in time. Furthermore, it is found that the model trained using the filtered data without a spanwise offset can well predict the large-scale variations at the target point when the spanwise offsets between the target point and the reference point are small (e.g., [Formula: see text] and [Formula: see text]). The proposed model leverages the power of neural networks and physical understanding. Further development of the model for complex scenarios will be carried out in the future work.