Abstract: Precise short-term forecasting of photovoltaic (PV) power is essential for grid stability and the integration of renewables. We propose two hybrid architectures—TS-BERT+ELM and PatchTST+ELM—separate temporal representation learning from regression by integrating transformer-based encoders with a ridge-regularized Extreme Learning Machine (ELM) for rapid, low-latency prediction. An evaluation of one-day-ahead predictions from 14-day input windows is conducted using daily PV datasets from five EU nations (Germany, France, Switzerland, Denmark, and the United Kingdom) provided from OPSD and enhanced with NASA POWER meteorological variables (global horizontal irradiance, cloud cover, and temperature) (f : $$ \mathbb{R}^{14×d} →\mathbb{R}$$). We present MAE, MSE, R2, and threshold accuracies (Accuracy@10%, Accuracy@50%), cexecute ablation, convergence, and sensitivity studies, and conduct paired t-tests and Wilcoxon signed-rank tests for statistical validation. Results indicate that TS-BERT+ELM regularly surpasses baselines on noisy and irregular datasets (France, Germany), whereas PatchTST+ELM demonstrates strong performance with high-quality, structured data (Denmark, UK); Switzerland occupies a position bridging the two categories. Integrating external weather-related features further enhances predictive accuracy and decreases variance, with statistically significant gains (p < 0.05) in four countries and an inconclusive UK case due to high variance. This modular design facilitates rapid convergence, maintains robustness against missing inputs, and enhances operational efficiency, and is compatible with federated and transfer learning for privacy-preserving, cross-site deployment. These findings support scalable, multimodal, and privacy-aware PV forecasting in real-world energy systems.

Keywords: Solar energy, TS-BERT, PatchTST, Extreme Learning Machine (ELM), hybrid deep learning, short-term forecasting

Pages: 316-335