A HOUSEHOLD ELECTRICITY CONSUMPTION PREDICTION AND ANOMALY DETECTION BASED ON FEATURES USING TWO-STEP UNSUPERVISED DEEP LEARNING APPROACH

Abstract

Accurate prediction of household electricity consumption is significant as it serves as a building block for effective energy management and operational decisions, essential for curtailing non-technical losses. A range of machine learning techniques have been implemented for detecting abnormal electricity consumption and have achieved significant results. However, with the evolution of anomalous electricity consumption coupled with the rapid growth in electricity consumption data, new challenges confronting anomalous electricity consumption are emerging. This current study proposes a two-step unsupervised machine learning approach, including a gated recurrent unit (GRU) regular network and a gated recurrent unit (GRU-autencoder) autoencoder, to detect consumption anomalies. The analysis was based on data collected from 450 households over a three-month period within the Tamale municipal assembly. The performance of the proposed model (the GRU autoencoder) was estimated using the MSE value, f-score, precision, and accuracy. The GRU autoencoder outperforms state-of-the-art methods in detecting anomalous electricity consumption, achieving an accuracy of 90.97% on the trained dataset.