Leveraging Anomaly Detection and AutoML for Modelling Residential Measurement Power Traces

Emerging data analytics approaches applied in power systems enable the extraction of relevant information from large scale measurement time series. These can be used to improve system observability and operation. Several open datasets can serve as benchmark that allow comparative results between pre-processing, modelling and evaluation strategies. We present a data analytics application on high reporting rate power measurements from single and multiple residential units. Outlier detection to label anomalies in the time series can be combined with automated machine learning libraries for efficient forecasting and dimensionality reduction. We apply this method for establishing an empirical relation between the occurence rate of outliers in the measurement time series and the forecasting performance. Results provide a discussion on the best, domain-adaptive, parametrisation and modelling options which are highly suited for power system measurements using open source software libraries.