@Article{ee.2026.085951,
AUTHOR = {Ye Tao, Qiong Cui, Haobo Chen, Jing Zhang, Zitong Tang, Zhuoying Liao, Jie Shu},
TITLE = {Photovoltaic Power Generation Prediction Based on the VMD-TCN-Transformer-GRU Model},
JOURNAL = {Energy Engineering},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/energy/online/detail/27495},
ISSN = {1546-0118},
ABSTRACT = {Photovoltaic (PV) output is significantly affected by meteorological factors such as irradiance, cloud cover, temperature, and humidity, and the power sequence typically exhibits strong non-stationarity and random fluctuations. To address the difficulty of a single model in simultaneously handling local variations and long-range dependencies, this paper proposes a hybrid prediction model based on Variational Mode Decomposition (VMD), Temporal Convolutional Network (TCN), Transformer, and Gated Recurrent Unit (GRU). First, key meteorological features are selected through Pearson correlation analysis, and similar day types such as sunny days and rainy days are classified using K-means clustering. Subsequently, VMD is applied to decompose the original power sequence into multiple relatively stationary subsequences. On this basis, TCN-Transformer-GRU prediction networks are constructed for each subsequence, in which TCN is used to extract local temporal features, Transformer is employed to characterize global dependency relationships, and GRU further enhances sequential memory capability. Finally, the prediction results of each subsequence are superimposed and reconstructed to obtain the final photovoltaic power prediction values. Validation results based on measured data from a photovoltaic power station in Guangzhou demonstrate that the proposed model achieves favorable performance across scenarios including high irradiation months, low irradiation months, sunny days, and rainy days. Compared with the single Transformer model, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are reduced by 83.7% and 84.3%, respectively, and coefficient of determination (R<sup>2</sup>) improves from 0.8585 to 0.9959. Cross-year test results also indicate that the model possesses good generalization capability. The research results can serve as a reference for photovoltaic power prediction, improvement of photovoltaic accommodation capacity, and subsequent research related to electricity-carbon coordination.},
DOI = {10.32604/ee.2026.085951}
}