A Similarity-Enhanced Transformer-LSTM Framework with IPOA for Short-Term Photovoltaic Power Forecasting

Abstract

Accurate prediction of PV output is critical for optimizing its absorption potential and ensuring the safe, stable, and cost-effective operation of the power grid. Yet, due to the intermittent and stochastic nature of photovoltaic power generation, establishing a highly precise prediction model presents significant difficulties. In this study, a hybrid forecasting framework integrating WCSD, CEEMDAN-FE, IPOA, and Transformer-LSTM is developed to improve PV power forecasting accuracy. Firstly, a new training data sample generation method based on WCSD is developed for determining the historical days having similar meteorological conditions to the predicted day. Secondly, CEEMDAN is employed to decompose original output sequence into an ensemble of components with different amplitudes and frequencies, where they were recombined as new set including a handful of components with low-frequency variation characteristics based on FE index. Thirdly, the IPOA is proposed for the first time, which couples Gaussian mutation and enhanced circle chaotic mapping. Next, the prediction model for each component is formulated by aid of Transformer-LSTM algorithm, the optimal hyperparameter combination of which is determined by IPOA method. Finally, the predicted results are obtained as the sum of individual component predictions. The prediction performance of the designed model is tested and verified via experimental analysis located in Yunnan Province, China and the publicly available Australian DKASC dataset. The empirical findings demonstrate that, in contrast to alternative benchmark models, our developed hybrid prediction model consistently attains superior prediction accuracy.