A Data-driven Detection Method Of Abnormal Consumption Behaviors For Restaurant Gas Users

Abnormal gas consumption behaviors of restaurant users including gas theft is a major concern commonly existing in the gas industry.Since abnormal gas consumption behaviors not only cause considerable revenue losses for gas supplier companies,but also endanger the public safety seriously due to the great risks brought to the pipeline network.Traditional detection methods of abnormal gas consumption behaviors highly rely on active human efforts,like full-scale on-site inspections,which are extremely costly while rather ineffective.Thanks to the wide deployment of smart meters,gas supplier companies are able to collect huge quantities of gas consumption data.Benefiting from that,we can devise a datadriven method to detect users suspicious of having abnormal gas consumption behaviors.In this paper,we propose an abnormal gas consumption behavior detection method for restaurant gas users,i.e.msRank,to discover suspicious restaurant users when only scarce historical labels are available.Our method contains three main components:(1)data pre-processing based on semantics,which filters reading noises in raw data,excludes data-missing users and zero-use users;(2)normal user modeling based on bi-seasonal mode clustering,which discovers typical seasonal gas consumption modes by clustering,generates bi-seasonal mode sequences based on them,then quantifies the self-stable seasonality of normal users by calculating the mode entropy,and distinguishes normal users from unstable ones;and(3)gas-theft suspect detection based on RankNet,which discovers gas-theft suspects among unstable users by RankNet-based suspicion scoring on extracted deviation features.By using detected normal users as negative samples to train RankNet,the component of normal user modeling and that of gas-theft suspect detection are seamlessly connected.In this way,we overcome the problem of label scarcity.We conduct extensive experiments on three real-world datasets provided by a Chinese gas supplier company.And the results demonstrate the advantages of our approach.We have developed and deployed an online system based on msRank,which provides a gas-theft suspect list weekly for the gas supplier company.