Enhanced Non-Intrusive Load Monitoring via Optimized Feature Selection and Majority-Voted Ensemble Learning

Precise recognition of different appliances using Non-Intrusive Load Monitoring (NILM) techniques is essential for numerous smart grid applications. Successful execution of NILM strategies depends on two factors: appropriate feature extraction followed by a suitable classification technique. Different appliance currents have different waveforms, which in turn have unique statistical features that can be exploited for performing appliance identification. This work proposes a multi-class statistical feature-assisted Ensemble Bagged Tree (EBT) approach for appliance recognition. Statistical features and impulsive metrics are extracted from the appliance currents. Prior to classification, feature selection is performed utilizing the minimal-redundancy-maximum-relevance (mRMR) method. The mRMR method assists in avoiding either the redundant features or features with less information, and thus only the most significant features are passed on to the classifier. Finally, for classification, the EBT method is utilized, which has superior performance, avoids over-fitting and has lower variance compared to single classifier based techniques. To validate the proposed approach, experimental appliance data from publicly available datasets were considered on both a host computer and an embedded device (Jetson Nano). The performance of the present work is also tested on various downsampled versions of public datasets.