Exploring the K-Nearest Neighbors Algorithm for Non-Intrusive Load Monitoring Using the COOLL Dataset

Abstract

In today’s technological era, electricity consumption is increasing as people rely more heavily on electrical appliances, resulting in higher electricity bills. Hence, managing household energy usage more effectively has become an important issue. A major challenge is that most users lack awareness of how to use electricity wisely and have limited visibility into the amount of electricity each appliance consumes. Non-Intrusive Load Monitoring (NILM) is seen as a promising solution, as it can disaggregate total energy usage into individual appliances without the need for additional sensors and allows users to monitor their electricity consumption more effectively. This study explores the K-Nearest Neighbors (kNN) algorithm and utilizes the COOLL (Controlled On/Off Load Library) dataset, which provides detailed appliance measurements for analysis. Three power-based feature sets, namely transient power, steady-state power, and a combined feature set are investigated. In addition, the K hyperparameter is varied, and classification accuracy is reported for both training and testing data. The results show that feature selection strongly affects kNN performance. The combined power feature set achieves the best overall performance, reaching a maximum test accuracy of 95.24% at K = 1, followed by steady-state power with a peak test accuracy of 90.48% (achieved at multiple K values), while transient power yields the lowest peak test accuracy of 76.79% at K = 15. Regarding the K value, transient features benefit from moderate K values, whereas the combined feature set performs best at small K, and overly large K values reduce accuracy due to averaging across too many neighbors. These findings show that combining transient and steady-state information provides a more discriminative representation for kNN-based NILM, and that appropriate K tuning further improves classification performance on the COOLL dataset.