Overhead Line Status Data Mining Based On ARIMA-LSTM

With the rapid development of the national economy and the growth of the power system, it’s important to ensure the safe operation of the power system and avoid outage loss. As an important component of the power system, the overhead line’s operation status along with the processing mode under abnormal condition is crucial to the operation of the power system.The power system presents the trend of digitalization, information and intelligence, with the large amount of the monitoring quantity for overhead line’s status uploaded to the data center. Meanwhile, the cooperation between power supply enterprises and meteorological departments is becoming more and more closely so that researchers can easily access the environmental factors of the impact of the overhead line. These status information from different sources have the characteristics of large scale, great variety and low value density, which constitute the large data of the overhead line’s status. Through analysis and mining of those data, we can accurately fit the change of the health indicators of the overhead line so as to improve the security and reliability of the power system.The paper proposes a model combined with time series analysis and deep learning. Firstly, the linear features of the key status are extracted by means of the auto regressive integrated moving average model. Secondly, using neural network in long short term memory unit(LSTM) to fit the residual of the model, the effects of external conditions on the device are taken, thus correct the time series model and improve the prediction accuracy. Using spark to make it run faster. Taking insulator contamination condition of 220 kV overhead line in Guangzhou as example, the time series model of equivalent salt density of insulator is established. Then, the micro meteorological elements around the tower calculated by spatial interpolation are used as the input of the LSTM neural network to estimate the residual of equivalent salt density. Results show that this method can accurately reflect the changes of the status in a period of time and can be run fast using and have a good speed-up ratio on Spark.