Attention-Enhanced Fusion Network for Precise Ultra-Short-Term PV Power Forecasting

Accurate photovoltaic (PV) power prediction is crucial for ensuring grid stability and optimizing energy dispatch. However, the intermittency of solar energy and the complex influence of weather conditions present considerable obstacles. While transformer-based methods excel at capturing long-range...

Full description

Saved in:
Bibliographic Details
Published in:IEEE Conference on Industrial Electronics and Applications (Online) pp. 1 - 6
Main Authors: Zheng, Dongyang, Zhu, Rongwu
Format: Conference Proceeding
Language:English
Published: IEEE 03.08.2025
Subjects:
Adaptation models
Benchmark testing
deep learning
energy conversion
feature fusion
Fluctuations
Forecasting
Long short term memory
Predictive models
Renewable energy sources
Solar energy
Stability analysis
Transformers
ISSN:2158-2297
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Accurate photovoltaic (PV) power prediction is crucial for ensuring grid stability and optimizing energy dispatch. However, the intermittency of solar energy and the complex influence of weather conditions present considerable obstacles. While transformer-based methods excel at capturing long-range dependencies, capturing fine-grained details, which is crucial for ultra-short-term forecasting, remains challenging. To address these limitations, this paper introduces a novel Attention- Enhanced Fusion Network (AEFN) for ultra-short-term PV forecasting. The AEFN model leverages a global multi-dimensional coordinate attention mechanism to adaptively capture spatial dependencies and refine feature representations. Furthermore, a hybrid LSTM-transformer network is employed to effectively model both short-term fluctuations and long-term trends, with LSTM focusing on capturing local temporal patterns and the transformer capturing global dependencies. Experimental results based on a real-world PV power dataset demonstrate that the AEFN model achieves superior performance compared to several benchmark forecasting methods.
ISSN:2158-2297
DOI:10.1109/ICIEA65512.2025.11149151