Home Appliances Detection Using U-net: Supervised Machine Learning Framework

Constant human intervention in tackling energy efficiency challenges is becoming a bit from the past. As moving forward to the future, smart grids are consistently improving through machine learning data analytics, from smart sensors to using algorithms to implement models that can classify and detect different power profiles without manual human efforts. Non Intrusive Load Monitoring (NILM) techniques are a leading technology in accessing the data from aggregated power signals of different facilities without installing meters on every appliance. This work aims to introduce one of the NILM techniques using a fully Convolutional Network model with the help of data already collected by a start-up based in Turin, Midori. The U-Net model, usually used for image segmentation, shows promise in this context due to its capability to be effectively trained with limited labelled datasets, closely mimicking real-world scenarios. Based on Midori datasets, the U-Net model is used for image segmentation. The framework can be trained without the need for huge annotated datasets but only with a few labelled images used in the training dataset, mimicking real-world problems. Using the limited collected data of appliance signatures, we train the model to predict the presence of these appliances among random power data and signatures, to lastly evaluate its performance over the testing set. The results obtained from the testing set indicate a high level of accuracy in appliance detection, which is a crucial step towards reducing energy consumption and promoting sustainability. Further research can explore the scalability of this model and its applicability in larger datasets.