Energy Disaggregation Using Binary Indicators for Appliance On/Off States

This study delves into energy disaggregation, a method to estimate individual appliance power consumption from aggregated building power data. Unconventionally, our approach employs binary indicators signifying appliance states during training, rather than the actual disaggregated data. We apply four preprocessing techniques Constant, Interval, Edge, and Markov to both aggregated power data and binary indicators to approximate disaggregated consumption. This approximation trains a denoising autoencoder, a neural network. We offer a comparison of preprocessing methods and the autoencoder's performance on data generated from them or true disaggregated data. Our goal is to assess whether the use of binary indicators significantly impacts neural network performance.