Information Extraction Approach for Energy Time Series Modelling

Increased adoption of energy monitoring devices across the energy system has resulted in large quantities of multivariate measurement data sets available for analysis at multi-scale resolutions. For buildings in particular, these can be leveraged to extract relevant information in order to characterize and improve its operation by establishing trends and anticipating faults before they occur. Several time series data mining algorithms have become available for efficient subsequence search and classification which can be adapted for domain-specific load profiling. We present an application of the Matrix Profile (MP) technique to energy time series for large commercial building load modelling. Several results are discussed which concern discord identification, building-specific MP values distributions as well as the effect of the particular distance metrics on the resulting processed input time series. Model free load forecasting can also serve as a suitable baseline for more advanced methods with contextual variables. Working with higher level information pieces leads to a speed up of the analysis and eliminates redundant raw data which makes the processed data suitable for online algorithm implementation and real-time building energy management.