Online incremental power forecasting method for centralized photovoltaic power plants based on deep learning

Under the "dual carbon" goals, centralized photovoltaic (PV) power stations have become a crucial support for the new energy power syst however, PV power generation is strongly intermittent and volatile due to factors such as seasons and weather. Addressing the issue of model performance degradation caused by the dynamic evolution of input data in practical acations, and the susceptibility of traditional updating methods to catastrophic forgetting, this paper proposes a deep learning-based online incremental power prediction model. The model introduces the Deep Experience Replay (DER) incremenng mechanism to construct a dual-core framework of "block feature extraction and online knowledge retention". It captures multi-scale periodic features via a patch token strategy, utilizes self-attention mechanto mine multivariate dependencies, and combines experience replay techniques to alleviate catastrophic forgetting. Experimental results based on real-world data from a PV power station indicate that the cumulative accuracy degradation rate of the prsed model is significantly lower than that of traditional models, demonstrating stronger adaptability and generalization capabilities, and providing an effective solution for online dynamic power prediction in centralized.