A New Simulated Database for Classification Comparison in Power Signature Analysis

The selection of the most appropriate detection, feature extraction and classification method is a fundamental step for the Non-Intrusive Load Monitoring (NILM) problem. In order to compare methods, a properly identified and annotated dataset is required. In this sense, several datasets have been proposed in the literature, real and simulated, with different features, loads and acquisition scenarios. In general, a common characteristic of these datasets is the absence of multiple simultaneous loads with a balance between the loads that are switched, precise indication of load events, and inclusion of noise and harmonic content. Such limitations may comprise a proper comparison between load disaggregation methods, hindering subsequent tasks, such as embedding the solution in electronic systems. With the aim of including all these requirements, this work presents a new simulated dataset using MATLAB-Simulink models, validated with real data, that controls the instant that each load is switched, allowing to precisely extract features during the transient of each load. Additionally, by varying the parameters of the simulation such as harmonic content and noise, it is possible to evaluate the performance of state-of-the-art methods (Voltage-Current Trajectories, Discrete Fourier and Wavelet Transforms) for load classification. In general, Voltage-Current Trajectory is the most affected method in low signal-to-noise ratio condition.