Akademik Veri Yönetim Sistemi
Arabian Journal for Science and Engineering, 2025 (SCI-Expanded, Scopus)
Accurate forecasting of solar energy production plays a critical role in ensuring the stability, reliability, and cost-effectiveness of renewable energy systems. This study proposes a hybrid forecasting framework that integrates advanced machine learning (ML) algorithms and a deep learning architecture, Temporal Fusion Transformer (TFT), to predict short-term solar power generation. A unique, high-resolution dataset is constructed by combining two years of hourly production data from 12 solar power plants across Turkey with detailed meteorological variables obtained from the Solcast platform. The dataset includes 30 environmental and operational features and is enriched with engineered variables such as cyclical time components, cloud factors, and performance ratios. Three categories of ML models such as regression-based, tree-based, and other ML models are systematically optimized using four different hyperparameter tuning strategies. Among these, XGBoost achieved the best performance with an R2 of 0.9620 and a mean absolute error (MAE) of 113.18 kWh. In parallel, the TFT model is customized for the solar forecasting task, incorporating attention mechanisms and quantile loss to improve both accuracy and interpretability. The TFT outperformed all other models, achieving an R2 of 0.9681 and demonstrating strong generalization across heterogeneous production sites. The results highlight the effectiveness of combining interpretable deep learning models with domain-specific feature engineering and optimization strategies. The proposed approach offers a scalable and practical solution for energy providers, contributing to smarter grid management, investment planning, and climate-aware energy policy development.