一种SH-ResNet模型的换流阀外冷却系统最优化选型方法

为在换流阀外冷却系统设计初期快速选择合理的冷却方式，分析了影响冷却方式选择的当地气候环境和换流站设备条件等因素，构建基于堆叠异构的残差网络(SH-ResNet)模型用于对冷却方式进行分类，模型集成了有监督的分类器与无监督的聚类方法，并将ResNet作为元分类器，深度挖掘输出结果之间的潜在联系。通过研究近年来不同地区站点的气候环境、换流阀冷却系统需求、设备布置情况以及最终设计使用的冷却方式，总计209个样本数据对所提出模型进行训练与评估试验。结果表明：SH-ResNet的分类正确率达到0.97,相较于基础分类器平均提高了11.46%,可见，在样本集较小的情况下该模型保证了其强大的泛化能力，并提高了分类准确度。基于该模型的冷却方式推荐系统交互窗口的设计不仅给予了冷却方式的推荐占比，并可视化特征参数与冷却方式间的联系，为换流阀外冷却系统设计提供了一种最优化选型方法。

Optimal selection method for external cooling system of converter valves based on SH-ResNet

For the efficient selection of a feasible cooling method before the design of external cooling system for a converter valve, the local climate environment and converter station equipment conditions were analyzed, and a residual network model based on stacking heterogeneous (SH-ResNet) was built to classify cooling methods. The model integrated supervised classifiers and unsupervised clustering algorithms, and ResNet was regarded as a meta-classifier to deeply explore the potential connections of the output results. The climate environment, converter valve cooling system requirements, equipment layout, and the corresponding cooling methods adopted in different regions in recent years were investigated, and a total of 209 samples were used to train and evaluate the proposed model. Results show that the classification accuracy rate of SH-ResNet reached 0.97, which was an average increase of 11.46% compared with the base classifiers. It indicates that the model maintains strong generalization ability and improves the classification accuracy in spite of small training sets. Finally, the interactive window of the cooling method recommendation system based on the proposed model was designed. It not only gives the recommended proportion for each cooling method, but also visualizes the relationship between the characteristic parameters and the cooling method, which provides an optimal method for the design of external cooling system of converter valves.