PGA–DSCformer–BiLSTM: A Hybrid Deep Learning Framework for Long-Term Load Forecasting
IEEE Access · 2026 · 14 · 92803-92820
DOI: 10.1109/access.2026.3704949 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Zeyu Chen und 3 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Jiangxi Provincial Key Laboratory of Intelligent Systems and Human-Machine Interaction, National College Students’ Innovation and Entrepreneurship Training Program of China.
Abstract (Originalsprache)
Electricity demand planning over long horizons requires forecasting models capable of capturing structural evolution and multi-scale temporal dynamics under increasingly complex load regimes. However, existing deep learning approaches often exhibit insufficient integration of localized pattern extraction and global dependency modeling, limited exploitation of bidirectional temporal correlations, and high sensitivity to manual hyperparameter configuration. To address these challenges, this study proposes a hybrid forecasting framework termed DSCformer–Bidirectional long short-term memory (BiLSTM) optimized by Phototropic Growth Algorithm (PGA). The proposed architecture enhances the Transformer encoder by embedding depthwise separable convolution to strengthen local temporal feature extraction, incorporates a bidirectional long short-term memory module to refine sequential state transitions, and employs the PGA for automated hyperparameter optimization. This layered design enables coordinated global–local representation learning, bidirectional contextual modeling, and adaptive parameter tuning. Extensive experiments on two real-world datasets demonstrate the effectiveness of the proposed approach. On the AUS dataset, the model reduces mean absolute error to 206.09 MW, achieving a 32.95% improvement over the strongest competing baseline. On the ENG dataset, mean absolute error decreases to 318.61 MW with a 38.74% improvement margin. The results confirm improved forecasting accuracy, stability, and cross-regional generalization for long-term load prediction.
Förderung
- National Natural Science Foundation of China
- Jiangxi Provincial Key Laboratory of Intelligent Systems and Human-Machine Interaction
- National College Students’ Innovation and Entrepreneurship Training Program of China
Zitieren
APA
Chen, Z., Tang, M., Liu, M., & Deng, S. (2026). PGA–DSCformer–BiLSTM: A Hybrid Deep Learning Framework for Long-Term Load Forecasting. IEEE Access, 14, 92803-92820. https://doi.org/10.1109/access.2026.3704949
Vancouver
Chen Z, Tang M, Liu M, Deng S. PGA–DSCformer–BiLSTM: A Hybrid Deep Learning Framework for Long-Term Load Forecasting. IEEE Access. 2026;14:92803-92820. doi:10.1109/access.2026.3704949
BibTeX
@article{Chen2026,
author = {Chen, Zeyu and Tang, Meng and Liu, Mingping and Deng, Suhui},
title = {PGA–DSCformer–BiLSTM: A Hybrid Deep Learning Framework for Long-Term Load Forecasting},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {92803--92820},
doi = {10.1109/access.2026.3704949}
}
EUVIMED