Outlier Analysis Of Urban Tail Transit Traction Energy Consumption Based On Machine Learning

Urban rail transit(URT)has become an essential public transportation for people’s daily life due to its large traffic volume,safety,efficiency and low pollution.With the rapid development of URT in recent years,the overall energy consumption has increased rapidly.Energy consumption of URT can be divided into train operation energy consumption,power and light energy consumption.The traction energy consumption during the train operation,accounts for the largest proportion of the total energy consumption,and it varies with influence factors such as passenger demand,lines,vehicles and control strategy.The investigation of abnormal conditions of traction energy consumption is an important means to discover the causes of energy consumption fluctuations,and the outliers in the data can provide quantitative basis for the investigation.At present,exceeding the fixed threshold can be defined as outliers in practical work,which doesn’t take the change of energy consumption pattern into account,and is prone to cause misreporting and underreporting.Considering the complexity of large URT and extensive data management,by analyzing the characteristics of its energy flow,influencing factors and energy consumption data,this thesis proposes an outlier analysis method based on machine learning.The main research contents are as follows.(1)Based on the existing energy consumption evaluation method of URT,two indicators of single train traction energy consumption and traction motors energy consumption are extended.An outlier analysis framework for single train and traction motors is designed according to the indicators.(2)A feature extraction method for traction energy consumption of URT is proposed.For single train traction energy feature data,random forest algorithm is used to select important features and reduce the data dimension;for traction motors,energy feature vectors are extracted from the traction energy consumption time series.Method verification is performed using data from a certain line of the Beijing Subway.The preprocessed data was used for subsequent outlier detection.(3)A method for detecting outliers of traction energy consumption of URT is proposed.For single train,a clustering algorithm is used to analyze energy consumption patterns,hypothetical test is used to verify their statistical significance,and a model-based outlier detection algorithm is used to identify outliers for each energy consumption pattern.For traction motors,clustering and classification algorithms are used to obtain the energy consumption pattern discrimination tree,and for each energy consumption pattern,a density-based outlier detection algorithm is used to identify outliers.The results show that there are three types of energy consumption patterns for single train energy consumption of a station and four types of energy consumption for traction units.Distinguishing energy consumption patterns can detect potential outliers compared with detecting outliers in the whole data set.(4)An analysis method of outliers’ degree of traction energy consumption of URT is proposed.The KNN-based scoring algorithm is proposed to obtain the degree of single train traction energy consumption outliers in order to improve the problem of misreporting and underreporting of outlier detection.Outliers with high scores need to be focused on.Finally,the train traction energy consumption evaluation software is designed and developed,and the proposed method is analyzed and verified.This thesis analyzes the outliers of single train traction energy consumption and traction motors energy consumption separately.The proposed method can dynamically adapt to changes in energy consumption patterns.The outliers can be located at the time and the stations.The outliers’ degree scoring improves the credibility.The theoretical methods and application examples have enriched the theoretical research of outlier detection of URT,and provided decision support for on-site outlier detection and energy conservation.