Real-Time Estimation of Total Solar Power Output in A Region Using An Efficient Hybrid Transformer - Quantile Regression - KDE Model

This study introduces a novel Transformer-Quantile Regression-Kernel Density Estimation model for real-time photovoltaic power estimation across Taiwan, aiming to improve the accuracy and timeliness of solar power forecasts for grid management and energy policy formulation. The complete framework integrates multisource data and advanced deep learning techniques, including MissForest for missing data imputation, to handle incomplete datasets effectively. To select the most representative PV sites, K-Means clustering is used, then optimized with the Silhouette Coefficient and combined with Pearson Correlation Coefficient. This clustering method allows for more efficient and accurate power generation estimation for the entire Taiwan region. Additionally, the model employs NAdam as the optimizer for training, ensuring better convergence and performance. The proposed model significantly improves prediction accuracy, providing probabilistic estimation through the QR-Transformer, which generates prediction intervals and enhances the practical utility and reliability of the forecasts. Comparative results show that this approach achieves the lowest error when compared with other models, such as BiLSTM, ANN, and GRU. The outcomes offer crucial data for energy managers, supporting resource allocation and risk assessment, and facilitating the development of solar and renewable energy policies and the optimization of energy strategies in Taiwan.