TFEformer: A new temporal frequency ensemble transformer for day-ahead photovoltaic power prediction

Abstract

The accurate prediction of day-ahead Photovoltaic (PV) power can provide technical support for complex solar management systems. This problem involves forecasting a long time series, and although several models have been proposed, the current state of stability and precision leaves much room for improvement. In this study, we introduce a novel Temporal Frequency Ensemble Transformer (TEFformer) designed for day-ahead PV power prediction, which integrates four crucial components: temporal attention, frequency attention, Fourier attention, and weather embedding. Initially, temporal attention is employed to directly model the original PV power series, providing fundamental insights. Fourier attention is devised to analyze the trend series derived from the original data in the frequency domain. Furthermore, frequency attention dissects seasonal variations into real and imaginary components for independent study, offering practical insights. Finally, numerical weather prediction data serves as future covariates, augmenting environmental information. Our experiments demonstrate that the TFEformer outperforms 9 advanced baseline models, showcasing its superior accuracy and robustness. Furthermore, in contrast to mainstream hybrid models that involve decomposition and frequency analysis in preprocessing, our proposed model exhibits higher efficiency.