A Physics Informed Machine Learning Approach for Non-intrusive Identification of Residential Photovoltaics

The development of real-time digital twins (DTs) of local area energy systems enables the integration of more distributed energy resources (DERs) into the electric grid and facilitates their efficient management. Digitalisation requires accurate recognition of DERs, which can be achieved by training machine learning (ML) models using a comprehensive labelled dataset that reflects energy penetration from DERs. However, acquiring such a dataset is challenging due to its intrusive nature. Therefore, this research study proposes the use of physics-informed machine learning (PIML) techniques for the non-intrusive identification of residences with photovoltaic (PV) power generation in a local area, using only the aggregated smart meter data of residences and weather data corresponding to the geographical location of the residences. In this paper, a physics-informed algorithm is used to label the smart meter dataset. The study compares the PV identification results obtained by a long short-term memory (LSTM), convolutional neural network (CNN), and Knearest neighbour (KNN) models using the proposed novel PIML methodology against those from a benchmark LSTM model. The test results show that the proposed physics informed methodology enables the LSTM model to achieve 81% performance accuracy in the context of non-intrusive PV identification.