Short-term Photovoltaic Forecasting: A Parallel TimesNet and AT-Informer-AT Method

As the integration of renewable energy accelerates, high accuracy Photovoltaic Power Generation Forecasting (PVGF) has become a key enabler for maintaining grid resilience and planning energy dispatch efficiently in distributed power network. To further improve forecasting performance and stability, we have developed a short-term PVGF method with a parallel architecture. Initially, Locally Weighted Scatterplot Smoothing (LOWESS) is applied to reduce data noise and stabilize the input sequences. Moreover, Feature Engineering (FE) is utilized to identify the most relevant input variables. Thirdly, a parallel model named ‘TNet-AIA’ is designed, which incorporates a parallel structure combining the strengths of TimesNet and Attention-Informer-Attention (AT-Informer-AT) models. Specifically, the TimesNet model is employed to capture multi-scale temporal patterns in the input sequences, while the AT-Informer-AT model successfully learns both long-term correlations and short-term local variations. Case studies are conducted on two representative Photovoltaic Power (PV) located in DKASC area, Alice Springs, Australia, and Xuhui District, Shanghai, China. Experimental findings indicate that the presented approach significantly improves the predictive performance and stability, achieving a notable 16.88% improvement in forecasting accuracy.