Multiscale Data Analytics for Residential Active Power Measurements through Time Series Data Mining

Abstract

Modern measurement and automation equipment for energy systems collect, store, process and communicate ever increasing quantities of raw data which can be used to build data-driven prediction and classification models. Directly using these large and unprocessed data sets can be inefficient especially in imbalanced class problems, where the positive class is sparsely represented in the training examples, such as classification of micro-scale transients. This is due to the longer time required for model training, and due to the increased possibility of obfuscating the useful information behind noisy readings. The matrix profile represents a computationally efficient and general purpose time series data mining technique which is suitable for embedded deployment in future generation smart meters, and in embedded energy gateways. Our analysis concerns the application of this technique on two types of residential electric power measurement data sets: a detached single family house and an apartment, with variable reporting rate and subsequence size parametrisation. Quantitative results support the findings that such approaches serve as practical instrument for measurement time series preprocessing in energy analytics.