基于人工智能的屋顶光伏资源评估方法及其应用

Method and Application of Rooftop Photovoltaic Resources Assessment Based on Artificial Intelligence

在我国加速推进整县分布式屋顶光伏的大背景下,为直观、可靠地初步评估建筑屋顶可利用的光伏资源,提出一种基于深度学习的屋顶光伏资源评估方法,将图像分割技术与光伏仿真部分结合并应用。先根据双卷积神经网络模型Double U-Net识别建筑屋顶的外轮廓,实现建筑物屋顶的高精度自动提取,进行屋顶边缘的检测及屋顶面积的计算,判断该屋顶光伏组件的数量,使用KLEIN-THEILACKER模型计算出在该屋顶光伏组件最佳倾角和其倾斜面辐照量,最终估算屋顶面积的光伏最佳发电量。提出的DoubleU-Net模型训练准确度可达95.83%,同时该方法可估算所选屋顶的月、年总辐照量及总发电量。最后,以河海大学常州校区为案例,通过与Solargis网站数据对比,验证了所提出方法具备较高的可靠性和适用性。

Under the background of accelerating the promotion of whole-county distributed rooftop photovoltaic(PV)in China,a deep learning-based rooftop PV resource assessment method which combined and applied image segmentation techniques with PV simulation components was proposed to assess the available PV resources on building rooftops intuitively and reliably.Firstly,the outer contour of the building roof was identified to realize the high-precision automatic extraction of the building roof according to the Double U-Net convolutional neural network model.Then the roof edge was detected,and the roof area was calculated to determine the number of photovoltaic modules on the roof.The KLEIN-THEILACKER model was used to calculate the optimal tilt angle of the PV modules and the irradiation on the titled surface.Finally the optimal photovoltaic power generation of the roof area was estimated.The training accuracy of the proposed Double U-Net model could reach 95.83%,and the proposed method could estimate the total monthly and annual irradiation and total power generation of the selected roof.Finally,the reliability and applicability of the proposed method were verified by comparing it with the data from the Solargis website,using the Changzhou campus of Hohai University as a case study.