Fault detection with hourly district energy data: Probabilistic methods and heuristics for automated detection and ranking of anomalies
2013 (English)Report (Other academic)
This project is motivated by the difficulties experienced by district energy utilities to detect faults in large-scale district energy systems. Faults that remain undetected can be costly and the industry loose credibility when customers detect faults and receive incorrect bills. Faults are common in district energy systems due to the high number of substations and instrumentation components. Also, the standard energy-metering instrumentation is designed for low cost and billing, not for automated fault detection. Large variations in building dynamics, building subsystems, human behaviour and the environment make the system complex to model and analyse. Therefore, conventional methods for fault detection are not applicable and the use of ad hoc methods for fault detection often result in numerous false alarms that are costly to analyse and manage. There is a growing interest among the utilities to develop services and functions that are based on data with high temporal resolution. Energy metering regulations are also expected to become more demanding in the future, which drives technology standards towards high-resolution data. This trend results in high rates of streaming data at the management level, which is more challenging to validate. Therefore, more efficient methods for fault detection are needed.This project deals with probabilistic methods for automated anomaly detection that are useful for the identification of faults in large-scale district energy systems. These methods are compatible with the information that is available in modern energy meter data management systems. We focus on methods and heuristics that can be applied automatically with a minimum of human assistance to enable cost-efficient analysis of data. With these methods, operators do not have to rely on ad hoc tests or manual inspection of graphs to detect anomalies in the data. Instead, operators can focus on the analysis of a subset of substations that are identified as abnormal. Intraday and intraweek variations in the thermal load are accounted for by automatically grouping hours of the week with similar thermal load characteristics. Alternatively, intraweek cycles can be accounted for by grouping days of the week with similar characteristics. Robust regression is used to model variable relationships with historical data. A robust outlier detection method is used to determine if variables deviate from the expectation defined by a regression model. Robust statistical methods are used to score outliers, so that outstanding substations can be identified automatically with a ranking procedure. The regression models can also be used for imputation of missing energy metering data, which is a common problem that is not always solved in an accurate way. We also present methods for the detection of long-term drift, which can be costly and otherwise difficult to detect, and the detection of poor precision in measurement data, which for example can result from oversized flow valves, misconfiguration and noise. In addition to fault detection, the proposed methods can be useful also for maintenance scheduling because substations that behave in a way that is consistent with the historical record can be given a lower maintenance priority compared to substations with abnormal behaviour.The proposed methods are studied using hourly data from a population of about one thousand district heating substations. Sample code of key functions is provided. We find that substations with documented faults, unknown faults and abnormal characteristics can be identified in about 5% of the substations. The lack of a well-defined dataset makes the development and evaluation of methods for fault detection challenging, and the fact that historical energy metering data includes abnormal data is often ignored in the literature. The proposed methods need to be implemented in a full-scale district energy management system under the supervision of experienced operators before the effects on the fault detection rate and cost efficiency can be properly evaluated. However, we are convinced that the proposed algorithms can be implemented in present data management systems and that they offer significant advantages over the methods that are commonly used today.
Place, publisher, year, edition, pages
Stockholm: Svensk Fjärrvärme , 2013. , 120 p.
Research subject Industrial Electronics
IdentifiersURN: urn:nbn:se:ltu:diva-22021Local ID: 12f0f9e8-d223-4442-88f8-f3c1af26f95dISBN: 978-91-7381-125-5 (PDF)OAI: oai:DiVA.org:ltu-22021DiVA: diva2:995069
Godkänd; 2013; 20131009 (fresan)2016-09-292016-09-29Bibliographically approved