基于改进Shapley特征优选及ResNet50深度学习模型的继电保护状态评估研究

Condition evaluation of relay protection system based on the modified Shapley feature optimization model and the ResNet50 deep learning algorithm

随着智能电网的发展,继电保护系统接入数据更为全面,而传统算法难以给出兼顾精度与效率的状态评估方法。提出了改进沙普利(Shapley)理论的特征优选模型,通过熵值法与Shapley理论综合分析各指标的重要性。构建了基于具有50个卷积层的残差网络(residual network-50,ResNet50)的继电保护状态评估的深度学习算法,介绍了算法的网络结构。选取上海某公司1796台继电保护装置的数据对算法进行测试,采用改进合成少数类过采样技术对不平衡数据进行处理,算法准确率高达96.7%,召回率达82.9%。与卷积神经网络和随机森林算法进行了对比分析,测试结果表明,ResNet50算法在准确率、召回率和精度方面都表现出明显的优势,可为现场巡检策略优化和故障排除提供支撑。

With the development of smart grids,the access data of relay protection systems are more comprehensive,and the traditional algorithms are difficult in condition evaluation with both accuracy and efficiency.The modified Shapley feature optimization model is proposed by comprehensively analyzing the feature importance using entropy method and Shapley method.A deep learning algorithm based on residual network-50(ResNet50)with 50 convolutional layers is constructed and the network structure of the algorithm is introduced.Using the data of 1796 relay protection devices in one electrical company in Shanghai,the proposed method is demonstrated.The imbalanced data is processed using the modified synthetic minority oversampling technique.The algorithm's accuracy rate is up to 96.7%and the recall rate is up to 82.9%.Compared to the convolutional neural network and random forest algorithms,the efficiency of ResNet50 is better especially in accuracy,recall,and precision,which can provide support for optimizing on-site inspection strategies and troubleshooting.