Natural gas consumption forecasting for anomaly detection. (15th November 2016)

• Record Type: Journal Article
• Authors: Baldacci, Lorenzo; Golfarelli, Matteo; Lombardi, Davide; Sami, Franco
• Is Part Of: Expert systems with applications. Volume 62(2016)
• Page Start: 190
• Page End: 201
• Publication Date: 2016-11-15
• Subjects: Anomaly detection -- Linear regression -- Local regression -- Gas flow forecast; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2016.06.013
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: An in-depth domain analysis is reported to derive reasonable modelling assumptions. We propose two original forecasting techniques for detecting anomalous gas flows. We developed an advanced approach to model the user behaviour to detect HDD. All the tests have been carried out on real data provided by a gas provider company. Abstract: Natural gas consumption forecasting is critical for many gas supplier companies tasks - e.g. gas procurement optimization, pipe network monitoring, management and security. This paper presents the joint work we carried out with HERA S.p.A., Italian gas provider leader, which goal is to forecast gas consumption for a given gas network as well as detecting anomalous gas flows according to historic data so to facilitate the monitoring and security processes in their central control room. Historic network conditions are sampled every 15 min, each sample is composed by a gas flow, an outside temperature, and the timestamp the sample was recorded. Descriptive analyses were carried out using historic data in a village and a small city, then two forecasting techniques were defined, one based on a nearest neighbor approach, one employing local regression analysis. Experimental results show that the historical data collected and stored can be used to reliably forecast gas consumption. A quantitative and qualitative comparison of the two methods is discussed in details so to highlight strengths and weaknesses. Moreover, due to the peculiarity of the domain, we worked with domain subject-matter experts to understand the capability of the methods in detecting anomalous gas consumption. Our results clearly show our forecasting techniques effectively support control room operators in identifying anomalous consumption. Providing a forecasting functionality is the first relevant step towards creating a full expert system that makes it easier for advanced operators to interpret the gas network behavior and that suggests the less-skilled ones the correct reactions to be taken upon the occurrence of anomalous events.