可再生能源2025，Vol.43Issue(7)：887-895,9.
基于多尺度卷积神经网络的屋顶光伏建筑轮廓提取方法研究
Building contour extraction algorithm for rooftop photovoltaic structures based on convolutional neural networks
摘要
Abstract
In promoting distributed photovoltaics across entire counties,assessing rooftop solar potential is crucial.Efficiently and accurately obtaining rooftop outlines in a given area is key to evaluating photovoltaic utilization potential there.In response to the advancement of county-wide distributed photovoltaic deployment in China,existing methods for extracting rooftop photovoltaic buildings have overlooked the multi-scale characteristics of images,leading to drawbacks such as blurred building outlines and low accuracy in extraction.This paper proposes a multi-scale enhanced convolution coupled with attention modulation for building contour extraction.Firstly,a multi-scale enhanced convolution module is constructed using dilated convolutions with multiple dilation rates,which is then integrated into the U-Net network to capture building features under different receptive fields,enabling the extraction results to more comprehensively express the overall and detailed features of buildings.Subsequently,the attention mechanism is introduced into the U-Net network to participate in skip connections,allowing for more precise extraction of building contours.Finally,a composite loss function is constructed using both cross-entropy loss and Dice coefficient loss functions to train the proposed model for building contour extraction.After integrating the irradiation assessment algorithm,the current rooftop can be assessed for irradiance based on latitude and longitude,as well as the tilt angle.Experimental results show that compared to current building contour extraction algorithms,the proposed algorithm not only has higher accuracy in building contour extraction but also performs well in extracting buildings of various scales.The proposed algorithm effectively improves the accuracy of building outline extraction,thereby enhancing the efficiency of assessing photovoltaic utilization potential.This method is valuable for advancing the use of GIS and artificial intelligence technologies in photovoltaic resource assessment,especially within the framework of county-wide promotion.关键词
屋顶光伏/建筑物轮廓提取/多尺度增强卷积模块/注意力机制Key words
rooftop photovoltaic/building contour extraction/multi-scale enhanced convolutional module/attention mechanism分类
能源科技引用本文复制引用
胡家宇,白建波,肖宇航,严家乐..基于多尺度卷积神经网络的屋顶光伏建筑轮廓提取方法研究[J].可再生能源,2025,43(7):887-895,9.基金项目
国家重点研发计划(2022YFB4201000). （2022YFB4201000）
