Machine Learning Algorithm for Assessing Photovoltaic Panels Partial Shading Losses based on Inverter Data

Partial shading is one type of fault where photovoltaic panels cast shadows between each other, reducing their production and hastening their ageing. In this paper, we document and describe two distinct Machine Learning models that aim to identify and assess the impact of partial shading in a real case study. These algorithms recognise similarities and patterns using expected and measured power data. The predicted power is calculated using the measured panel irradiance, current, and voltage using a photovoltaic panel electric circuit model. The first Machine Learning model employs K-means clustering to analyse the differences between expected and measured power, grouping data based on these deviations. The second Machine Learning model leverages the outputs of the K-means model as labels for a Long Short-Term Memory neural network, which classifies periods of partial shading. Experimental data from both models are presented, with the K-means model achieving a closer approximation to the reference values. However, the Long Short-Term Memory model demonstrated flexibility and scalability without requiring prior dataset knowledge from the end user.