Two-stage photovoltaic power forecasting method with an optimized transformer

Abstract

Accurate photovoltaic (PV) power forecasting ensures the stability and reliability of power systems. To address the complex characteristics of nonlinearity, volatility, and periodicity, a novel two-stage PV forecasting method based on an optimized transformer architecture is proposed. In the first stage, an inverted transformer backbone was utilized to consider the multivariate correlation of the PV power series and capture its non-linearity and volatility. ProbSparse attention was introduced to reduce high-memory occupation and solve computational overload issues. In the second stage, a weighted series decomposition module was proposed to extract the periodicity of the PV power series, and the final forecasting results were obtained through additive reconstruction. Experiments on two public datasets showed that the proposed forecasting method has high accuracy, robustness, and computational efficiency. Its RMSE improved by 31.23% compared with that of a traditional transformer, and its MSE improved by 12.57% compared with that of a baseline model.