传统的换流站数字化建设三维建模方法已经存在很多种，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，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，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。文中基于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，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，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5 s内实现，峰值信噪比（peak signal-to-noise ratio, PSNR）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 traditional 3D modeling methods for digital construction of converter stations, and most of the modeling algorithms for converter stations cannot achieve real-time real scene modeling, making it difficult to achieve real-time fault monitoring based on 3D modeling. At the same time, the artificial intelligence image recognition algorithm of the converter station is difficult to implement during the training process due to the need for large-scale datasets. The data collection process will consume a large amount of energy and human resources, increasing energy consumption. This paper generates a three-dimensional model of converter station equipment using small-scale inspection images based on Instant-ngp, the MobileNetV3-small model is improved to obtain the mobilenet for smart grid (MSG) model, which further reduces the sample size required for training, and the 3D results of the model generated by Instant-ngp are used to capture images and input them into the MSG model for training, solving the problem of training set size. The experimental results show that the modeling of converter station equipment using Instant-ngp can be achieved within 5 seconds, and the peak signal-to-noise ratio (PSNR) index can reach 20-25dB. Using the 3D modeling results of Instant-ngp, under the same hardware conditions, the average inference speed of MSG compared to original model of MobileNetV3-small has increased by 7.3%, and it has reached over 85% in different power system scenarios, which greatly reduces the energy consumption required for the digital construction of converter stations.