Multi-Head Attention Based Model for Non-Intrusive Appliance State Detection in Smart Buildings

Non-intrusive load monitoring, also known as energy disaggregation, is the task of identifying the status of the electrical appliances and their energy consumption in a building from the knowledge of the aggregate energy consumption alone. Recently, many deep learning approaches have successfully handled this problem. Generally, the deep learning models for the task take the low-frequency aggregate energy data collected by the smart meter of the building and estimate the status of the target appliance (ON/OFF). In this paper, an attempt is made to further improve the performance and generalizability property of the task by adopting the powerful multi-head attention mechanism from the field of language translation. In addition, two parallel branches of the convolutional layer are also placed to extract the features of the input at different scales, which assists the attention mechanism in improving the performance. The model's performance is compared with another deep learning-based state-of-the-art model using two publicly available datasets. The test results establish the superior performance and generalizability of the proposed model.