基于区域自由的地基云图云种类识别方法  

Ground Based Cloud Recognition with an Anchor Free Method

作　　者：李逸超 郭睿 张少迪 寿泽锋 陈静 刘翼飞 LI Yichao;GUO Rui;ZHANG Shaodi;SHOU Zefeng;CHEN Jing;LIU Yifei(State Grid Shanghai Electric Power Company Pudong Power Supply Company,Shanghai 200000,China;Shanghai Key Laboratory of Smart Grad Demand Response,Shanghai Electrical Apparatus Research Institute(Group)Co.Ltd.,Shanghai 200000,China;College of Automation Engineering,Shanghai University of Electric Power,Shanghai 200090,China)

机构地区：[1]国家电网上海市电力公司浦东供电公司,上海200000 [2]上海电器科学研究所(集团)有限公司上海市智能电网需求响应重点实验室,上海200000 [3]上海电力大学自动化学院,上海200000

出　　处：《哈尔滨理工大学学报》2023年第2期128-135,共8页Journal of Harbin University of Science and Technology

基　　金：国网上海电力公司2021年营销专项(6409212000ND);国家自然科学基金青年基金(51607111)。

摘　　要：准确识别云种类有助于提高预测光伏发电功率精度。针对地基云图云种类识别目标候选框选择复杂、识别速度慢等问题,提出了基于区域自由的地基云图云种类识别方法。首先以Center Net为地基云图云种类识别的基本架构,通过热力图预测,关键点预测、中心点预测和候选框预测构建了区域自由的地基云图云种类识别流程。然后,设计了云种类识别模型的主干网络、损失函数和候选框预测方法。最后,以CenterNet-Resdcn101为模型,从算法识别精度、候选框预测置信度和识别速度等方面与主流目标识别方法和云种类识别方法进行了分析比较。实验结果表明本文所提方法具有更高的识别精度和更快的识别速度。Accurate recognition of cloud type is helpful to improve the forecasting accuracy of photovoltaic power production.In order to solve the problems of cloud type recognition of the ground-based cloud such as complex target candidate box selection and slow detection speed,a recognition method of cloud types in ground-based cloud map based on anchor free is proposed.First,the paper takes Center Net as the basic architecture of cloud type recognition.Based on thermodynamic diagrams prediction,key point prediction,center point prediction and candidate box prediction,a anchor free ground-based cloud type detection process is constructed.And then,the main network,loss function and candidate box prediction method of cloud type recognition model are designed.Finally,take CenterNet-Resdcn101 as the model,compared the algorithm recognition accuracy,candidate boxes predict confidence and identify speed with mainstream target recognition methods and cloud type recognition method.The results showed that the cloud type recognition method of the paper has higher recognition accuracy and faster recognition speed.

关 键 词：云分类 云种类识别 区域自由 Center Net 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]