Research on early warning of wire icing risk level based on machine learning

Abstract

This paper based on the daily wire icing observation data at the observatory from 1980 to 2018, selected the main influencing factors via the relationship between wire icing and meteorological elements by statistical analysis. With uses a variety of machine learning methods to establish a wire icing prediction model. Finally, take the snowfall process on February 14,2020 was evaluated to assess the forecast accuracy of the model. The results show that the icing on electric wires by the observatory mainly occurs from December to January of the following year, which of them are dominated rime icing type. The correlation between daily minimum and maximum temperature and icing thickness is better than that of daily average temperature. With the increase of freezing thickness, the daily minimum temperature and relative humidity are more concentrated at 0% and 700%. The nonlinear prediction model can better simulate the design ice thickness of the wire ice accretion. Based on the existing training set, using random forest and neural network, the accuracy of the algorithm is significantly higher than that of linear regression and support vector machine, and the prediction model still has the phenomenon of high value underestimation. Multi-layer perceptron predicted the best results, with less intense icing in areas with significant snowfall.