MEED: An unsupervised multi-environment event detector for non-intrusive load monitoring

The accurate detection of transitions between appliance states in electrical signals is the fundamental step that numerous energy conserving applications, such as Non-Intrusive Load Monitoring, rely on. So far, domain experts define rules and patterns to detect changes of appliance states and to extract detailed consumption information of individual appliances subsequently. Such event detectors are specifically designed for certain environments and need to be tediously adapted for new ones, as they require in-depth expert knowledge of the environment. To overcome this limitation, we propose a new unsupervised, multi-environment event detector, called MEED, that is based on a bidirectional recurrent denoising autoencoder. The performance of MEED is evaluated by comparing it to two state-of-the-art algorithms on two publicly available datasets from different environments. The results show that MEED improves the current state of the art and outperforms the reference algorithms on a residential (BLUED) and an office environment (BLOND) dataset while being trained and used fully unsupervised in the heterogeneous environments.