NILM Data-Set for Varying Operating Voltages

Globally building sector energy consumption is increasing rapidly. Improving building energy efficiency is essential for sustainability. Load monitoring provides detailed consumption feedback to enable consumers to save energy. Non-Intrusive Load Monitoring (NILM) is a cost-effective way to identify individual appliance energy consumption from aggregate energy consumption. Machine learning-based NILM techniques have been proposed in the literature to accurately identify appliances.In the real-world, the operating conditions may vary from training conditions affecting the performance of supervised NILM techniques. The performance of supervised NILM techniques degrades in presence of data variations due to noise, sensor drift, or source voltage fluctuations. This study evaluates the performance of supervised event detection and appliance identification techniques under varying operating voltages. An automated setup captures the aggregate consumption data of various home appliances. We employ steady state-supervised event detection and appliance identification using standard learning algorithms such as K-NN, Naive Bayes, Decision Tree, and Random Forest.The results show that varying voltage data significantly affects the performance of classifiers. We also evaluate mitigation strategies such as normalization, feature selection, and class balancing for improving classifiers. Insights from the experimental results help in developing robust NILM systems that can overcome the effects of data variations.