CTformer: Transformer-Based Scenario Generation with Complementary Meteorological Inputs

Amid growing reliance on weather-dependent photovoltaics (PV), generating high-fidelity scenarios that accurately represent uncertainty has become a cornerstone of reliable grid planning and operation. This paper proposes CTformer, a novel Transformer-based framework that fuses PV output and cloud cover trajectories as epistemically complementary inputs. Unlike conventional models that rely on uni-source statistics or autoregressive decoding, CTformer introduces a complementary-aware encoder that explicitly captures causal asymmetry and semantic divergence across modalities, followed by a non-autoregressive generative decoder that maps noise to structured future trajectories. The architecture integrates physical intuition with machine learning flexibility, enabling the generation of diverse, statistically coherent, and temporally faithful scenarios. Extensive evaluations on real-world datasets demonstrate that CTformer outperforms both neural and ensemble baselines in replicating quantile accuracy, probabilistic calibration, and power spectral structure, offering a scalable and interpretable solution for scenario-based PV forecasting under deep uncertainty.