Identifying Distinct Performance Patterns in Utility-Scale Photovoltaic Plants Using an Unsupervised Machine Learning Model

Photovoltaic (PV) systems have achieved an impressive cumulative capacity of 1.2 TW in 2022, emerging as a pivotal component in energy production. However, this significant adoption also highlights the crucial need for accurate estimation of PV system power generation, for gauging electricity availability, assessing investment feasibility, and determining profit margins. Although a large amount of data is collected from operational PV plants, it is challenging to determine their performance, as electrical parameters need to be normalised to account for environmental conditions — a procedure that is often associated with significant uncertainty. The concept of a digital twin, a virtual replica of the PV plant, is a powerful method to determine the expected behaviour, and therefore, can be used to detect deviations in performance or identify faults. However, the reliance on digital twins based upon physical models poses challenges, as they require many input parameters (some with significant associated uncertainties) and certain simplifications or assumptions (such as linear degradation or soiling), thereby, imposing limitations on achievable accuracy and predictive capability. This study proposes a novel method to overcome these limitations by leveraging machine learning (ML) algorithms to create learnt digital twins of PV plants, using only the collected raw daily operational data. Hence, the creation of these digital twins does not require the imposition of any physical models with underlying assumptions or biases. The developed data-driven approach is demonstrated using operational data from a utility-scale bifacial PV plant. We show that the developed method automatically learns to represent latent factors within the raw operational data, enabling the identification of different behavioural patterns, such as operation under clear-sky or cloudy conditions. It also correctly infers correlations between electrical and meteorological features, and uncovers implicit temporal characteristics in daily performance, without being provided with this explicit knowledge or information. This novel ML-based digital twin methodology marks a critical step toward completely automated data-driven Operations and Maintenance (O&M) systems that are based on real-world operational behaviour, with the ability to continuously update and improve over time. This will broadly enhance the reliability of utility-scale PV, supporting the much-needed transition to renewable energy.