Reconstruction of flow fields around large-span flexible PV array based on physics-informed deep learning

The complex three-dimensional inter-row flow interference induced by bidirectional series of large-span flexible PV arrays is the primary reason for the wind-induced damage. However, wind tunnel tests are difficult to capture the flow fields subjected to inter-row interference. By comparison, deep learning methods offer promising solutions for accurate reconstruction and prediction of complex flow fields. The present study focuses on the 5-row by 3-span PV array with a span of 40 meters, located at the State Power Investment Group Flexible PV Demonstration Base in Yancheng, Jiangsu. Large-eddy simulations of wind fields were performed to provide training data for a fully connected neural network deep learning method that incorporates the loss function into Navier-Stokes equations. Based on this innovative method, a data-driven model and a data-physics-driven model are established for accurately reconstructing the velocity and pressure fields around an array of large-span flexible PV. Compared to the data-driven model, the data-physics-driven model demonstrates superior efficacy in capturing the flow characteristics. Specifically, the streamwise velocity reconstruction errors upstream of the first two rows and the fourth row, at the upper and lower edges of the PV panels and in the wake region are reduced by 60.2% and 36.6%, respectively. The reconstruction errors of the lateral velocity at the upper and lower edges of the PV panels are reduced by 53.7%. Moreover, the model has full-field reconstruction errors of 16.6% and 18.5% for the streamwise and lateral velocities, respectively. Meanwhile, when the pressure term is not considered in the loss function, the data-driven model cannot learn pressure information from the training data, while the data-physics-driven model can obtain the pressure field through the velocity field, yielding an average reconstruction error of the inter-row pressure fields being only 16.1%. This study provides a reference for a new intelligent reconstruction method for the flow fields around complex structures under wind loads.