基于Instant-ngp的低碳目标换流站实时建模与样本生成方法研究

江一; 乔柱桥; 王玉俊; 李章浦; 张凯; 梅红伟

江一,乔柱桥,王玉俊,李章浦,张凯,梅红伟.基于Instant-ngp的低碳目标换流站实时建模与样本生成方法研究[J].电测与仪表,2025,62(12):55-62.
jiangyi,qiaozhuqiao,wangyujun,lizhangpu,zhangkai,meihongwei.Research on Real Time Modeling and Sample Generation Method for Low Carbon Target Converter Stations Based on Instant-ngp[J].Electrical Measurement & Instrumentation,2025,62(12):55-62.

基于Instant-ngp的低碳目标换流站实时建模与样本生成方法研究

Research on Real Time Modeling and Sample Generation Method for Low Carbon Target Converter Stations Based on Instant-ngp

DOI：

中文关键词: 换流站数字化建设三维建模算法深度学习故障检测

英文关键词:Converter station, Digital construction, 3D modeling algorithm, Deep learning, Fault detection.

基金项目:南方电网公司科技项目（基金编号：010900KK52200010）

作者中文名	作者英文名	单位
江一	jiangyi	中国南方电网有限公司超高压输电公司昆明局
乔柱桥	qiaozhuqiao	中国南方电网有限公司超高压输电公司昆明局
王玉俊	wangyujun	中国南方电网有限公司超高压输电公司昆明局
李章浦	lizhangpu	清华大学深圳国际研究生院
张凯	zhangkai	清华大学深圳国际研究生院
梅红伟	meihongwei	清华大学深圳国际研究生院

摘要点击次数: 9

中文摘要:

传统的换流站数字化建设三维建模方法已经存在很多种，代表性的建模方法主要有BIM（Building Information Modeling）建模和激光扫描建模。但是因为技术的自身局限，大部分换流站的建模算法都无法做到实时的真实场景建模，因此很难实现基于三维建模的故障实时监测。同时，换流站的人工智能图像识别算法在训练过程中因为需要大规模的数据集而落地困难。换流站在数字化建设以及数据采集过程中，会消耗大量的能源以及人力资源，增加能源消耗。本文首先基于Instant-ngp使用较小规模的巡检图像生成换流站设备的三维模型。然后对MobileNetV3-small模型进行改进得到MSG（MobileNet for Smart Grid）模型，再次缩小训练所需样本规模。最后，使用Instant-ngp生成的模型三维结果截取图像输入MSG模型进行训练，解决训练集规模问题。实验结果显示，Instant-ngp对换流站设备的建模可以在5s内实现，PSNR(Peak Signal-to-Noise Ratio)指标可以达到20-25dB，利用Instant-ngp的三维建模结果，在相同硬件条件下，MSG相较于MobileNetV3-small原始模型推理速度平均提升了7.3%，在不同电力系统场景下都达到了85%以上，大大减少了换流站数字化建设所需要的能源消耗。

英文摘要:

There are many types of three -dimensional modeling methods for digital construction of traditional power system converter station. The representative modeling methods mainly include BIM (Building Information Modeling) modeling and laser scanning modeling. However, because of the limitations of technology, most of the modeling algorithms of stations cannot achieve real -time real scene modeling, so it is difficult to achieve real -time monitoring of three -dimensional model -based faults. At the same time, the artificial intelligence image recognition algorithm at the converter station is difficult to land due to the lack of large -scale data sets during training. In the process of digital construction and data collection, converter stations will consume a large amount of energy and human resources, increasing the total amount of energy consumption. This article first uses a smaller-scale inspection image to generate a three-dimensional model based on Instant-ngp. Then improve the MobileNetV3-small model to get MSG (MobileNet for Smart Grid) model, and reducing the sample scale required for training. Finally, we cut the image from the three-dimensional result generated by the model generated by Instant-NGP, and enter the picture into the MSG model to solve the problem of training sets. The experimental results show that the modeling of Instant-NGP on the converter station can be implemented within 5S, and the PSNR(Peak Signal-to-Noise Ratio PSNR) indicator can reach 20-25dB. Using Instant-ngp""s three-dimensional modeling results. Under the same hardware conditions, MSG reasoning speed increased by 7.3% average compared with MobileNetV3-small, reaching more than 85% in different power system scenarios. The results of this article greatly reduce the energy consumption required for digital construction of converter stations.

查看全文查看/发表评论下载PDF阅读器

关闭