This paper presents the Down Sampling Feature Extractor (DSFE), a novel and extendable algorithm for Non-Intrusive Load Monitoring (NILM). The proposed algorithm addresses the dual challenges of sampling frequency reduction for electrical appliance load signatures and the automatic extraction of discriminative features from the down-sampled data. Its extendable nature is characterized by two key hyper-parameters: the down-sampling factor and the dimensionality of the output feature vector. The DSFE algorithm is implemented via an autoencoder neural network, which learns feature representations end-to-end directly from NILM data. These high-quality features enable the use of simple, computationally efficient classifiers, such as k-Nearest Neighbors (KNN), for the load classification task. Evaluated on the WHITED dataset, the proposed method achieves a load classification accuracy of 95.66%, surpassing the state-of-the-art result of 88.59%. Crucially, this performance gain is achieved with a model containing 125 × fewer parameters, demonstrating that the DSFE algorithm significantly enhances both the accuracy and computational efficiency of NILM-based load classification. • A Novel, Efficiency-Focused Framework : We introduce an autoencoder-based framework that performs intelligent down-sampling and discriminative feature learning end-to-end. This innovation directly tackles the challenge of processing high-frequency electrical data by reducing its dimensionality without losing critical information for appliance identification. • Substantial Performance Gains with Drastic Model Simplification : Evaluated on the public WHITED dataset, the DSFE algorithm achieves a state-of-the-art load classification accuracy of 95.66%. Crucially, this superior performance is accomplished with a model that contains 125 × fewer parameters than the previous best-performing approach. This remarkable reduction in complexity is a key step toward embedding efficient NILM models in resource-constrained edge devices. • Practical Implications for Energy Systems : By enabling high-accuracy classification with a simple, lightweight classifier (k-NN), our work significantly lowers the computational barrier for NILM. This advances the practical viability of NILM for applications such as real-time energy disaggregation, fault detection, and demand-side management, which are central themes in Energy and AI.
Brito et al. (Fri,) studied this question.