Enabling Alarm-Based Fault Prediction for Smart Meters in District Heating Systems: A Danish Case Study

Abstract

District heating companies utilize smart meters that generate alarms that indicate faults in their sensors and installations. If these alarms are not tended to, the data cannot be trusted, and the applications that utilize them will not perform properly. Currently, smart meter data are mostly used for billing, and the district heating company is obligated to ensure the data quality. Here, retrospective correction of data is possible using the alarms; however, identification of sensor problems earlier can help improve the data quality. This paper is undertaken in collaboration with a district heating company in which not all of these alarms are tended to. This is due to various barriers and misconceptions. A shift in perspective must happen, both to utilize the current alarms more efficiently and to permit the incorporation of predictive capabilities of alarms to enable smart solutions in the future and improve data quality now. This paper proposes a prediction framework for one of the alarms in the customer installation. The framework can predict sensor faults to a high degree with a precision of 88% and a true positive rate of 79% over a prediction horizon of 24 h. The framework uses a modified definition of an alarm and was tested using a selection of machine learning methods with the optimization of hyperparameters and an investigation into prediction horizons. To the best of our knowledge, this is the first instance of such a methodology.