Ultra-Short-Term Photovoltaic Power Prediction Based on Multi-Timescale Feature Optimization and Transformer

Introduction: To enhance multi-timescale feature extraction and mitigate noise interference in PV power forecasting, this study presents an ultra-short-term PV power prediction method that incorporates Transformer-optimized multi-timescale feature integration. Methods: The PV power series is decomposed using the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) and reconstructed into high-, medium-, and low-frequency components based on the Zero Crossing Rate (ZCR) and sample entropy (SE). The Extended Kalman Filter (EKF) is used to denoise high- and medium-frequency components in a dynamic manner. WTConv1d combines wavelet and convolution operations to extract multi-scale features. The processed features and meteorological data are fed into a Transformer model to capture dependencies and forecast power output. Results: Tests on real data from PV stations in Xinjiang and Gansu, China, show the model achieves high accuracy and robustness. At Xinjiang, eMAE is 0.761, eRMSE is 1.426, and R2 is 0.991. At Gansu, the values are 0.693, 1.309, and 0.988 respectively, confirming the model’s effectiveness across different regions. Discussion: The proposed model significantly improves forecasting accuracy and robustness, especially in multi-step predictions. It addresses limitations in existing models regarding feature loss and noise interference. However, challenges remain in generalization and real-time deployment. conclusion: In this paper, an efficient feature extraction method for multi-timescale feature optimization is proposed, and its effectiveness is verified in real ultra-short-term PV power prediction. The proposed method fully combines the advantages of data decomposition reconstruction denoising and WTConv1d, which ensures that the features of different time scales are fully extracted while avoiding the problem of information loss caused by traditional data decomposition-based PV power prediction methods. Conclusion: This study presents an efficient multi-timescale feature extraction method validated on real PV data. By combining ICEEMDAN, EKF, and WTConv1d, the model fully extracts time-scale features while minimizing information loss from traditional decomposition approaches.