Model design for photovoltaic facilities based on fuzzy neural network as core of its digital twin

This study presents the development of the core of a digital twin for a photovoltaic (PV) facility located at CIESOL-Almería. Two modeling approaches are proposed: a physics-based model using an equivalent electrical circuit, and a data-driven neurofuzzy model based on an adaptive neuro-fuzzy inference system (ANFIS). The neurofuzzy model, designed as a gray-box system, offers high interpretability and adaptability, and stands out for its rapid synchronization capability with the physical asset, enabling real-time behavior modeling essential to the digital twin framework. The ANFIS-based model accurately captures the dynamic power output of the PV system and is suitable for integration into energy management strategies based on predictive modeling. The model exhibits strong predictive performance, with a worst-case mean absolute error of only 16.37 W, a standard deviation of 126.22 W, a standard error of 0.73 W, and a coefficient of determination of 0.99, indicating high consistency and accuracy. When compared to the equivalent electrical circuit model and a previously published artificial neural network applied to a lower-capacity PV system, the neurofuzzy model demonstrates superior accuracy. Specifically, the normalized mean absolute error and normalized root mean square error are 0.0036 and 0.0282 per watt, respectively, outperforming both the equivalent electrical circuit model (0.01982 and 0.0520 per watt) and the neural network approach. These differences represent relative improvements of 87.39 % and 61.55 % over the neural network benchmark. In addition, the neurofuzzy model requires significantly lower computational resources, making it suitable for real-time applications and implementation in industrial controllers. The results confirm the potential of gray-box neurofuzzy modeling as a core component of a digital twin, providing a reliable, efficient, and interpretable foundation for control, monitoring, and optimization of PV installations. • Neurofuzzy and physics-based models form the core of a PV digital twin. • The Neurofuzzy model enables fast twinning in the digital twin framework. • Twinning uses 87 days of 1-min PV data and accounts for seasonal variations. • The Neurofuzzy model is accurate, precise, and computationally efficient. • The model is interpretable and suitable for predictive control in real time.