Photovoltaic power forecasting: Using wavelet threshold denoising combined with VMD
Lin Liu, Jianqiu Zhang, Shibei Xue
Abstract
In photovoltaic power generation systems, the power output is highly influenced by weather changes. The significant changes in weather patterns and frequencies result in significant fluctuations in regional power generation, posing serious challenges to the safe and stable operation of power systems. Compared to the ignorance of the practical existence of features in existing works when employing signal processing methods for decomposition and feature construction, this paper proposes an innovative deep-learning-based algorithm that integrates wavelet threshold denoising with Variational Mode Decomposition techniques to enhance the accuracy of PV power prediction through feature construction. The method proposed in this paper applies VMD decomposition to photovoltaic signals processed by wavelet threshold denoising to obtain IMFs. By selecting and predicting IMFs, it constructs usable IMF features for photovoltaic prediction. The experimental results demonstrate that, after incorporating the proposed IMFs features, the MAE and SMAPE of the models are reduced by an average of 12.68% and 17.65%, respectively. These results fully validate the significant enhancement of prediction performance by the constructed IMFs features. This study provides a practical and widely applicable solution for PV power prediction, effectively addressing the limitations of traditional methods in real-world applications. • VMD-based PV prediction often assumes IMF usage, limiting real-world application. • PV data become smoother after wavelet denoising, aiding experiments and use. • IMFs are selected after VMD and predicted to generate features for modeling. • Incorporating IMF features improves PV prediction, reducing MAE by 12.68%. • Transformers outperform others via self-attention for key feature extraction.