Improving Long-Term Photovoltaic Degradation Forecasting with Hybrid Neural Networks

Abstract

Photovoltaic (PV) cells degrade over time due to various environmental factors. Determining the rate of such degradation is important for solar operators as it informs the maintenance and replacement cycle of their hardware. Degradation impacts modules the most soon after installation, so early detection is necessary. Since not all failure modes are detectable through physical inspection, it may be necessary to infer the degradation pattern by analyzing the module’s power output over time. While previous works have focused on long-term PV generation capacity forecasting, they rely only on trends in production data, and do not take into account the underlying physical processes behind the degradation. Instead, we propose identifying the degradation mode through a random forest classifier and using this information along with a state-space model forecast to train a variational autoencoder that generates a prediction of future energy production along with a confidence bound. We compare our proposed model with existing state-space and neural approaches and find that our hybrid neural network approach outperforms the baseline approaches in terms of Winkler score, a forecasting metric that balances precision and confidence interval width, on the Desert Knowledge Australia Solar Centre dataset.