基于时空图卷积网络的输电线路覆冰预测

Transmission Line Icing Prediction Based on Spatio-temporal Graph Convolutional Networks

关键词:  覆冰预测|数据探索|时空分布特征|图卷积网络|时空预测

Key words:  icing prediction|data exploration|spatio-temporal distribution characteristics|graph convolutional network|spatio-temporal prediction

摘要:

针对已有的输电线路覆冰预测模型鲜有考虑覆冰过程中的空间特征信息，从而导致预测精度欠佳的问题，本文从时空序列预测的角度建立输电线路覆冰方面的预测体系，采用图卷积网络(Graph Convolutional Network,GCN)构建输电线路覆冰预测模型，基于图神经网络设计对输电线路覆冰拉力的图数据进行深度特征学习与图特征向量表示，以更好地提取电网塔杆覆冰拉力值的时空分布特征，从而准确预测未来的拉力值。基于南方电网的真实实验数据，设计一套可靠的数据预处理流程，将电网覆冰拉力数据转化为可以深度学习的时空序列大数据进行训练和验证。实验结果表明，本文提出的模型较已有的主流覆冰预测模型具有更加优异和稳定的预测结果，能够为输电线路及时除冰工作提供决策参考。

Abstract:

In view of the fact that the existing transmission line icing prediction model rarely considers the spatial feature information in the icing process,resulting in poor prediction accuracy,a prediction system for transmission line icing from the perspective of spatio-temporal series prediction was established in this paper,and used the Graph Convolutional Network (GCN) to establish a transmission line icing prediction model.Based on the graph neural network design,the deep feature learning and graph feature vector representation of the graph data of the icing tension of the transmission line were carried out to better extract the spatial and temporal distribution characteristics of the icing tension value of the power grid tower,so as to accurately predict the future tension value.Based on the real experimental data of China Southern Power Grid,a set of reliable data preprocessing process was designed and implemented to transform the icing tension data of the power grid into spatio-temporal sequence big data that could be deeply learned for training and verification.The experimental results show that the model proposed in this paper has more excellent and stable prediction results than the existing mainstream icing prediction model,which can provide a decision-making reference for the timely deicing work of the transmission line.