STL-Twinsformer-TTAO: A short-term photovoltaic power forecasting model based on seasonal-trend interactive learning

Accurate photovoltaic (PV) power forecasting is critical for ensuring grid stability yet remains impeded by the inherent non-stationarity and uncertainty of solar generation. To address these challenges, this paper proposes the STL-Twinsformer-TTAO framework, a novel model based on seasonal-trend interactive learning. The methodology first employs Seasonal-Trend decomposition using Loess (STL) to decouple raw power series into stationary trend, seasonal, and residual components. These features are subsequently processed by the Twinsformer's interactive dual-stream architecture, where the learning of seasonal patterns actively guides the representation of trend evolution to capture latent coupling relationships. Additionally, the Triangular Topology Aggregation Optimizer (TTAO) is introduced to efficiently determine optimal hyperparameters. Comprehensive evaluations using real-world data from an Australian PV station indicate that the proposed model surpasses eight established baselines, achieving an average Normalized Mean Squared Error (nMSE) reduction of 14.9% across four seasons. Furthermore, robustness experiments at a distinct site, Yulara, reveal a substantial average 31.5% nMSE reduction compared to the baseline model. These findings highlight the model's robustness against spatial heterogeneity, suggesting promising potential for generalization across varying micro-meteorological conditions.