Accurate long-horizon photovoltaic (PV) power forecasting remains constrained by the absence of reliable future meteorological data and the limited adaptability of existing models. To address these challenges, this study proposes an end-to-end forecasting workflow that integrates year-ahead meteorological data synthesis, optimisation-driven model tuning, and hybrid deep-learning architecture design into a single coherent process. The framework employs a Conditional Time Generative Adversarial Network (CTimeGAN) in combination with a Large-Language-Model-based module to generate physically consistent, trend-aware meteorological sequences that embed plausible future variations while preserving statistical consistency. To improve predictive performance, an Enhanced Alpha Evolution Optimiser (EAEO) featuring diversity-adaptive exploration and rank-weighted updating is employed to tune a Transformer–LSTM hybrid network that jointly captures global dependencies and temporal dynamics. Comparative analyses show that the proposed framework consistently achieves superior performance across all evaluated hybrid architectures, demonstrating improved accuracy and robustness in long-term forecasting. Compared to a baseline LSTM model, the proposed method reduces MAE and RMSE by approximately 40–46%, while R 2 increases by around 38%. The framework provides a scalable, privacy-aware solution for high-resolution, year-ahead PV power forecasting, supporting energy storage scheduling, demand response optimisation, and the reliable integration of distributed renewable resources in future net-zero power systems. • A unified framework enables high-resolution year-ahead solar power forecasting. • Year-ahead weather scenarios are generated using data-driven trend refinement. • Adaptive optimisation improves deep learning model robustness and stability. • Long-range temporal patterns are effectively captured for annual power prediction. • Annual forecasting errors are reduced by over 40% using real historical data.
Qi et al. (Fri,) studied this question.