Explainable deep learning models for smart meter data

Deep learning (DL) plays an important role in developing various sectors with a growing demand of data. DL’s techniques have been developed to tackle a particular task in energy sectors, energy disaggregation or non-intrusive load monitoring (NILM). Total energy consumption can be separated into individual patterns for each appliance using data from a smart meter. In this paper, popular sequence-to-point (seq2point) and WaveNets are used for energy disaggregation. In addition, explainable techniques are used as integrated gradient (IG), Deep Learning Important FeaTures (DeepLIFT), and occlusion. Those attributions are applied for disaggregation model to explain how algorithm decides on disaggregation output corresponding with input features by heat map visualization. Explainable models are essential for DL’s methods because they help non-expert users understand DL’s decision. Deep learning possibly has ability to handle energy disaggregation and explainable methods, especially occlusion. IG and DeepLIFT sensitively focus on specific input features.