ABSTRACT Operational optimisation and predictive maintenance are essential for upgraded reliability and cost‐efficiency in renewable energy systems. This study proposed an integrated model combining physics‐informed neural networks (PINNs), adaptive decision‐making and transfer learning to improve the performance of solar farms and wind turbines. The model leverages domain knowledge to capture precise fault identification, enlarge failure prediction lead times and enable rapid deployment across geographically diverse installations with nominal site‐specific data. Field evaluations across multiple solar and wind installations illustrate that the PINN‐based framework achieves up to 87.3% accuracy in component fault prediction with lead times of 14.2 days for blade issues and 21.3 days for gearbox failures, outperforming commercial condition monitoring systems and conventional machine learning. Innovative algorithms to optimise the cleaning of solar panels have shown improved energy production (8.3%), reduced water consumption (31.2%) and decreased labour requirements (34.1%). The architecture that has been used for the edge‐computing systems supports analytics in real time and has had an impact on maintaining operational capabilities above 93.2% even with disruption in communication. This forecast also indicates that through the use of the model's ability to perform transfer learning; it provides the opportunity to capture more than 85% of the initial model's performance when installing on new installations with only 65.1% of the necessary initial training samples while overcoming the cold‐start challenge. Although there are limitations due to the reliance on detailed component specifications, environmental variability, and the length of time required for system adaptation, the results have demonstrated significant economic and operational advantages. The practical and scalable implementation of this concept will allow for the continued implementation of Predictive and Resource‐Efficient Renewable Energy Operations.
Muhamad et al. (Thu,) studied this question.