Signal-Guided Dual-Predictor Learning for Ultra-Short-Term PV Power Forecasting

Abstract

To address the insufficient adaptability of a unified predictor caused by the non-stationarity of photovoltaic (PV) power series, pronounced multi-scale fluctuations, and differences in the predictable structures of decomposed subcomponents, this paper proposes an ultra-short-term PV power forecasting method based on CEEMDAN-VMD and signal-property-guided dual predictors. First, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is employed to decompose the original power series, and the resulting Intrinsic Mode Functions (IMFs) are reconstructed into high-, medium-, and low-frequency sequences using K-means clustering. Subsequently, Variational Mode Decomposition (VMD) is applied only to the high-frequency component for secondary decomposition. Furthermore, Sample Entropy (SE), Central Frequency (CF), Energy Proportion (EP), First-order Autocorrelation Coefficient (Lag1ACF), and Fluctuation Index (FI) are used to characterize the properties of the final components, based on which differentiated modeling is performed. Specifically, H1, H2, and Low are modeled using the Regularity Predictor (RP), whereas H3–H6 and Mid are modeled using the Dynamic Predictor (DP), which integrates a multi-scale dilated Temporal Convolutional Network (TCN) with a Gated Recurrent Unit (GRU). The forecasting results of all components are directly summed to obtain the final PV power forecast. Experiments conducted on seasonal datasets from a PV power station in Northwest China showed that, compared with several representative deep learning models, the proposed method reduced the Mean Absolute Error (MAE) by 72.49%, 61.54%, 61.79%, and 44.10% in spring, summer, autumn, and winter, respectively, and reduced the Root Mean Square Error (RMSE) by 79.91%, 68.61%, 77.33%, and 63.16%, respectively, with the coefficient of determination (R²) ranging from 0.9948 to 0.9991. The Wilcoxon test indicated that the performance improvements were statistically significant. On the independent Alice Springs dataset in Australia, the MAE, RMSE, and coefficient of determination (R²) were 0.0793, 0.1308, and 0.9988, respectively, further validating the generalization capability of the model. The results demonstrate that component-level differentiated forecasting based on signal properties can effectively improve the accuracy and model adaptability of ultra-short-term PV power forecasting.