Dual temporal feature-enhanced stacked ensemble learning for hourly PV energy prediction: A case study on three sites in lisbon, Portugal

Photovoltaics (PV) has emerged as a leading renewable energy source; however, its inherently unstable output, particularly at high levels of grid penetration, poses a substantial challenge. While deep learning (DL)-based forecasting models have been explored, they require extensive data and computational resources, limiting their practical application. Additionally, most existing methods represent time categorically, which fails to capture sequential dependencies accurately. To cope up with these limitations, this study proposes a stacked ensemble model named the Stack model to forecast hourly PV-produced energy for three PV systems in Lisbon, Portugal. The Stack model integrates multiple tree-based methods, incorporates Random Forest regressor, Extreme Gradient Boosting, and Light Gradient-Boosting Machine as base learners, with linear regression as the meta-learner to enhance accuracy. Unlike previous approaches, our model incorporates explicit numerical and cyclic encoding of time, improving the ability to learn temporal dependencies. The Stack model outperformed baseline methods across all sites validated on real-time data, including Joao (RMSE: 0.7, R2: 0.9608), Julia (RMSE: 1.64, R2: 0.9555), and Zarco (RMSE: 1.5, R2: 0.9472). A two-step verification statistical validation process confirmed the model's robustness and reliability. These findings highlight the Stack model’s effectiveness in overcoming computational and time representation challenges over the previous studies offering an efficient solution for PV energy forecasting.