Unsupervised machine learning for photovoltaic systems: Modelling LID dynamics in SHJ solar cells

Abstract

We present an unsupervised machine learning-based modelling and analysis procedure for the study of high-dimensional non-linear dynamic systems using experimental characterisation data, herein used to investigate the dynamics of solar cell performance changes arising from a light-induced degradation (LID) mechanism observed in silicon heterojunction (SHJ) solar cells during illuminated annealing at elevated temperatures. A training dataset of state trajectories (tensors through time) was prepared from solar cell characterisation results (changes in cell performance during prolonged exposure to various combinations of temperature and illumination intensity). A generative representation learning architecture was developed, featuring neural network (NN) models to map between state trajectories and time-independent latent-states, and a latent ordinary differential equation model parametrised by a deep NN to predict latent-state transition dynamics. These models were trained together as a β-variational autoencoder to ensure highly disentangled latent-state variables. Inference of the trained model indicates good accuracy in LID dynamics prediction concurrently across all state variables. The trained model successfully identified the primary relationship between illumination intensity (and not temperature) and the kinetics of degradation and recovery, in agreement with previously reported analysis performed using traditional methods. However, this was achieved here in a completely unsupervised manner in under 2 minutes training time on consumer-grade hardware, and can be similarly applied for investigations into a broad range of photovoltaic systems.