Enhanced multi-horizon photovoltaic power forecasting: A novel approach integrating ICEEMDAN decomposition with hierarchical frequency neural networks

Abstract

As a crucial renewable energy source, solar PV power generation drives environmental protection and energy transformation. However, existing forecasting models struggle to accurately capture the complex dynamics of photovoltaic (PV) power, primarily due to monolithic modeling paradigms and inadequate representation of temporal information. To address these challenges, this paper proposes a novel hybrid model that leverages data decomposition and frequency-stratified prediction. The model employs the advanced ICEEMDAN algorithm to address complex non-stationarity. Additionally, it introduces a frequency-stratified heterogeneous network for precise component-wise modeling and integrates Improved Relative Positional Encoding (IRPE) to accurately capture temporal dependencies. To comprehensively evaluate model performance, this study employs quantile regression to generate probabilistic prediction intervals, using the median output as the baseline for point predictions. The model’s performance is validated through ablation experiments and comparisons of single-step and multi-step predictions with recent benchmark models. The results indicate that the model excels under the MIMO strategy, achieving normalized nMAE values of 0.1142 and 0.1490 for 120-minute and 2880-minute forecasts on the DKASC and Solar I datasets, respectively, surpassing recent baseline models by 14.6% and 8.1%. Furthermore, the model’s statistical stability and robustness are confirmed through 30 independent Wilcoxon signed-rank tests, as well as an uncertainty analysis conducted under various weather conditions. In summary, the model’s high accuracy and stability provide robust support for power plant operations and planning.