基于深度学习与稀疏风洞数据的长跨柔性光伏结构时空风压场预测

Abstract Long-span flexible photovoltaic (PV) structures are one of the key solutions to the challenges of the “PV+” development. However, their long span, light weight, flexible stiffness, and high clearance result in pronounced wind-induced vibration responses, categorizing them as wind-sensitive structures. Consequently, wind load becomes the controlling load in their structural design. Currently, there is no established standard for determining wind loads on flexible photovoltaic structures, and research in this area primarily relies on wind tunnel testing. However, due to limitations in scale, wind pressure measurement points can only be sparsely distributed on the structure’s surface. This study integrates wind tunnel test data with deep learning methods to predict wind pressure spatiotemporal fields on flexible photovoltaic structures based on a limited number of monitoring points. Specifically, considering the spatiotemporal fluctuation characteristics of wind pressure, a fully convolutional network model (FCN) with multi-scale and skip connections based on a CNN framework is proposed. This model includes multiple pathways, enabling it to capture wind pressure information across various vortex structural scales of the spatiotemporal wind pressure field. Consequently, it can predict the complete spatiotemporal wind pressure field using data from a limited number of measurement points. The results show that, in the spatial dimension, the relative error and overall error between the predicted and ground truth are within 10%. In the temporal dimension, the predicted time-history curves closely align with the ground truth, with a match rate of over 90% for fluctuating responses. This indicates that the FCN model can effectively learn the spatiotemporal characteristics of wind pressure on the surface of long-span flexible photovoltaic structures. Therefore, the proposed method for predicting wind pressure spatiotemporal fields on long-span flexible photovoltaic structures offers significant potential for optimizing the spatial arrangement of wind pressure sensors and predicting wind pressure time histories, holding significant scientific and engineering application value.