Identification of representative energy consumption patterns: development and comparative analysis of methodologies

Abstract

This paper presents the development and empirical evaluation of a methodology for identifying the most representative daily energy consumption pattern, a key task for optimizing Smart Grid systems, load forecasting, and demand management. The proposed approach focuses on optimizing data preprocessing and utilizing a multi-factor similarity metric. It includes a two-stage adaptive filtering of daily profiles based on peak and average power, which allows for the effective exclusion of anomalous and inactive days through dynamic thresholds calculated based on statistical indicators (median, percentiles), adapting to individual consumption characteristics. To select the representative pattern, a composite metric was developed that comprehensively integrates similarities in shape, power level, and the temporal characteristics of active periods. The scientific novelty lies in the hybrid filtering and direct medoid selection without full-fledged clustering, which simplifies computation and eliminates the need to predefine the number of clusters. The methodology provides a fast and justified determination of the most typical consumption profile, which can serve as a reliable baseline model for anomaly detection, consumer segmentation, or synthetic load data generation. Testing results on real data from the "ECO data set" confirm the approach's effectiveness for various electrical appliances. This research makes a significant contribution to the advancement of energy management methods. Prospects include the integration of "elastic" metrics, formalized clustering, and the development of sophisticated anomaly detection systems.