HiFAKES: High-frequency synthetic appliance signatures generator for non-intrusive load monitoring

원문 초록

Non-intrusive load monitoring (NILM) rely on data-driven methods and require extensive datasets of power consumption over a long period of time. However, the scarcity of datasets negatively impacts the performance of current NILM algorithms. Further, while the vast majority of these datasets are made by collecting low-frequency sampling power consumption patterns, high-frequency datasets, which would significantly improve the accuracy of load identification, are comparatively few due to the time and costs that such data collection entail. As a response to these problems, synthetic datasets have been developed. Existing synthetic datasets for NILM, while useful, typically depend on large training datasets and are evaluated using a single aggregated test metric, which cannot precisely reflect the fidelity, diversity, and novelty of the data. To address this gap, we introduce HiFAKES, a novel physics-informed, training-free generative model designed to produce high-frequency sampling synthetic power signatures. HiFAKES generates synthetic data with superior fidelity and diversity. Indeed, the $\alpha$-precision and $\beta$-recall indicators are respectively about 2.5 times larger than those for existing models, this makes our proposed model highly effective for NILM applications. HiFAKES operates by generating reference signatures templates based on real appliance signatures data, and by replicating these with some features variations, thus simulating both known and novel appliances. A key advantage of this model is its ability to generate unlimited high-frequency synthetic appliance data from a single training example. The generated data closely mimics real-world power consumption, as demonstrated by comprehensive evaluations, including the so-called 3D metric ($\alpha$-precision, $\beta$-recall, authenticity) and domain classifier tests.