Deep Neural Networks for Appliance Transient Classification

Smart plugs are useful devices for measuring the appliance load, but intrusive. It has long been the goal of energy companies and researchers to monitor the load of all household appliances in a non-intrusive manner, using only a single smart meter. We show that deep neural networks can be extremely effective in this regard. Automatic feature learning can pick out distinctive load-dependent shapes in the time series power data, enabling identification of appliance signatures. We find that properly arranged deep neural networks are capable of multi-class appliance classification, outperforming a traditional multiclass classification algorithm. We evaluate on the public PLAID dataset, and compare results with features extracted from sampling frequencies in the 1Hz-1kHz range. We show that adding features extracted from high frequency sampling significantly improves classification performance over data obtained at typical smart meter frequencies.