Education, Science, Technology, Innovation and Life
Open Access
Sign In

Day-Ahead Photovoltaic Power Forecasting Based on Pseudo-Label Contrastive Learning and Weather-Conditioned Gated Fusion of Physical Bases

Download as PDF

DOI: 10.23977/acss.2026.100220 | Downloads: 2 | Views: 113

Author(s)

Bihui Lv 1, Fang Liu 1, Xuri Fan 1, Wei Chen 1

Affiliation(s)

1 China Huadian Overseas Investment Co., Ltd., Beijing, China

Corresponding Author

Bihui Lv

ABSTRACT

To address biases in numerical weather prediction (NWP) for day-ahead photovoltaic power forecasting and the lack of physical constraints in purely data-driven end-to-end regression, which can produce nonzero nighttime output or power exceeding physical limits, this paper proposes a forecasting method based on weather pseudo-label contrastive learning and gated fusion of multiple physical bases. First, a clear-sky power benchmark is established for each station using solar geometry and a clear-sky irradiance model. A clear-sky index is constructed from historical power measurements to generate weather pseudo-labels, and supervised contrastive pretraining is used to enhance the weather representation learned by the historical encoder. Three physical bases are then constructed from NWP irradiance, the clear-sky index at the same time on the previous day, and the most recent valid clear-sky index. A gating network performs a time-step-wise convex combination of these bases according to weather conditions, with Softmax-normalized weights ensuring nonnegative predictions, compliance with capacity limits, and zero nighttime output by construction. Experiments on ten stations in the PVOD dataset show that the proposed method generally outperforms multiple baselines in MAE, RMSE, and R². Ablation experiments verify the contribution of each physical basis, while cross-station adaptation experiments further demonstrate that parameter-efficient fine-tuning achieves accuracy close to full fine-tuning at a substantially lower training cost.

KEYWORDS

Day-Ahead Photovoltaic Power Forecasting; Weather Pseudo-Labels; Supervised Contrastive Learning; Physical Bases; Gated Fusion

CITE THIS PAPER

Bihui Lv, Fang Liu, Xuri Fan, Wei Chen. Day-Ahead Photovoltaic Power Forecasting Based on Pseudo-Label Contrastive Learning and Weather-Conditioned Gated Fusion of Physical Bases. Advances in Computer, Signals and Systems (2026). Vol. 10, No. 2, 175-185. DOI: http://dx.doi.org/10.23977/acss.2026.100220.

REFERENCES

[1] Antonanzas, J., Osorio, N., Escobar, R., et al. (2016) Review of Photovoltaic Power Forecasting. Solar Energy, 136, 78-111.
[2] Hochreiter, S. and Schmidhuber, J. (1997) Long Short-Term Memory. Neural Computation, 9(8), 1735-1780.
[3] Zeng, A., Chen, M., Zhang, L. and Xu, Q. (2023) Are Transformers Effective for Time Series Forecasting? Proceedings of the AAAI Conference on Artificial Intelligence, 37(9), 11121-11128.
[4] Wu, H., Hu, T., Liu, Y., et al. (2023) TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. International Conference on Learning Representations.
[5] Holmgren, W.F., Hansen, C.W. and Mikofski, M.A. (2018) pvlib python: A Python Package for Modeling Solar Energy Systems. Journal of Open Source Software, 3(29), 884.
[6] Khosla, P., Teterwak, P., Wang, C., et al. (2020) Supervised Contrastive Learning. Advances in Neural Information Processing Systems, 33, 18661-18673.
[7] Vaswani, A., Shazeer, N., Parmar, N., et al. (2017) Attention Is All You Need. Advances in Neural Information Processing Systems, 5998-6008.
[8] Houlsby, N., Giurgiu, A., Jastrzebski, S., et al. (2019) Parameter-Efficient Transfer Learning for NLP. Proceedings of the 36th International Conference on Machine Learning, 2790-2799.
[9] Hu, E.J., Shen, Y., Wallis, P., et al. (2022) LoRA: Low-Rank Adaptation of Large Language Models. International Conference on Learning Representations.
[10] Yao, T., Wang, J., Wu, H., et al. (2021) A Photovoltaic Power Output Dataset: Multi-Source Photovoltaic Power Output Dataset with Python Toolkit. Solar Energy, 230, 122-130.

Downloads: 51800
Visits: 1378627

Sponsors, Associates, and Links


All published work is licensed under a Creative Commons Attribution 4.0 International License.

Copyright © 2016 - 2031 Clausius Scientific Press Inc. All Rights Reserved.