AI-Driven Digital Twin Framework for Long-Term Photovoltaic Integration in Smart Cities: An Eleven-Year Comparative Study of Silicon Photovoltaic Technologies Under Semi-Arid Climate Conditions

Abstract

The reliable long-term prediction of photovoltaic (PV) system performance is essential for optimizing operation, maintenance, and energy management in smart cities. Digital Twin (DT) technology has emerged as a promising approach for real-time monitoring and predictive analytics by continuously integrating physical system measurements with virtual models. However, many existing DT-based studies rely on limited validation periods, lack comprehensive uncertainty quantification, and provide insufficient comparisons with conventional forecasting approaches. To address these limitations, this study proposes a data-driven Digital Twin framework for daily photovoltaic energy production prediction for three silicon photovoltaic technologies (amorphous silicon, polycrystalline silicon, and monocrystalline silicon) operating under semi-arid climatic conditions in Morocco. The framework integrates data quality assessment and statistical production trend analysis based on linear regression, bootstrap confidence intervals, the Mann–Kendall trend test, and Sen’s slope estimator. In addition, the proposed DT model was benchmarked against persistence, linear regression, Random Forest, XGBoost, LightGBM, and Long Short-Term Memory (LSTM) models using a chronological training, validation, and independent testing framework. Model performance was evaluated over an independent test period of 602 consecutive days, representing approximately 1.65 years of continuous operation and more than one complete annual cycle. The proposed Digital Twin improved upon the standalone LightGBM model, achieving an RMSE of 1.3047 kWh/day, an MAE of 0.9255 kWh/day, a MAPE of 14.62%, and an R2 of 0.7015, compared with an RMSE of 1.3349 kWh/day, an MAE of 0.9637 kWh/day, and an R2 of 0.6875 for LightGBM. Over the 11-year monitoring period, the estimated long-term production trend rates were −0.566% yr−1 for a-Si, −0.335% yr−1 for pc-Si, and −0.260% yr−1 for mc-Si. Bootstrap analysis yielded median long-term production trend rates of −0.573, −0.326, and −0.251% yr−1, respectively, while Mann–Kendall tests indicated no statistically significant monotonic production trend for any technology (p = 0.1611, 0.2758, and 0.3502, respectively). The extended independent testing period, combined with statistical production trend analysis and comparative machine-learning evaluation, demonstrates the applicability of the proposed Digital Twin framework for photovoltaic performance monitoring and adaptive prediction under semi-arid climatic conditions.