Ultra-short-term photovoltaic power forecasting based on CEEMDAN-VMD and BiTGRU-ACGN fusion

Abstract

Photovoltaic (PV) power forecasting is of great significance for ensuring the safe and stable operation of power systems. In response to the nonlinearity of PV power time series and their complex local and global temporal correlations, this study proposes an ultra-short-term forecasting framework based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)-variational mode decomposition (VMD) and bidirectional temporal convolutional gated recurrent unit (BiTGRU)-attention-based convolutional gating network (ACGN). First, the original PV time series is subjected to an initial decomposition using CEEMDAN. K-means clustering is then performed based on sample entropy and center frequency, with a secondary VMD decomposition applied only to the high-frequency cluster. Subsequently, this study constructs a parallel network comprising a BiTGRU and an ACGN: the first branch integrates bidirectional dilated causal convolutions with gated recurrent units to collaboratively capture local dynamic features; the second branch integrates positional encoding, a multi-head self-attention mechanism and a gated convolutional feedforward neural network to model global temporal correlations. The synergistic fusion of features from both branches simultaneously enhances the model’s ability to capture both local details and global patterns, while Adaptive Bandwidth Kernel Density Estimation is used to generate prediction intervals at multiple confidence levels. Experiments were conducted using data from a power station in China and the Alice PV dataset in Australia. The results demonstrate that the proposed model exhibits excellent predictive performance across different seasons. In particular, in summer, R2 improved by 0.03%–3.16%, MAE decreased by 13.60%–74.06%, and RMSE decreased by 13.14%–82.13%. Furthermore, the results of the Wilcoxon test results were all statistically significant, confirming that this method possesses excellent predictive performance and generalization capability, and can effectively support the optimized dispatching and practical operation and maintenance of smart grids.