TransKANs-Fusion: 融合时空稀疏注意力与KAN 的光伏预测方法

打开文本图片集

中图分类号：TM615 文献标识码：A 文章编号：1672-0792（2026）02-0069-10

DOI： 10.3969/j.ISSN.1672-0792.2026.02.008

TransKANs-Fusion： Photovoltaic Prediction Method via Integrating Spatio-temporal Sparse Attention and KAN

ZHOU Ziguan1， ZHU Yaping1， LIU Zhu1 ， HUANG Chao1， CAO Junwei²， TU Guoyu²， HUANG Yan2，3

1.Beijing State Grid Electric Power Technology Co.，LTD.，Beijing 100176，China

2.NationalResearch Center for InformationScience and Technology，Tsinghua University，Beijing10o084， China

3.Department of Automation， Tsinghua University，Beijing 1ooo84， China

Abstract： To address the inherent intermittency and uncertainty of PV power generation， under the global transition towards clean energy， this paper proposes a photovoltaic power prediction method via integrating spatio-temporal sparse atention and KAN （TransKANs-Fusion）. This framework integrates a newly designed encoder based on Kolmogorov-Arnold networks （KAN），alocal-global sparse attention mechanism，and a multikernel one-dimensional convolution module， enabling efective capture of the spatiotemporal dynamic pattrns in time-series data.The method combines the strengths of physics-informed regularization and data-driven approaches，enhancing its capability to characterize PV power generation under complex meteorological conditions.The proposed TransKANs-Fusion framework achieves highly accurate and stable forecasting performance while maintaining comparable computational efficiency.

Key words： photovoltaic prediction; sparse attention; KAN; 1D convolution; temporal feature

0引言

光伏发电固有的强间歇性与气象依赖性，导致其功率输出呈现高度非线性、时空耦合及动态不确定性，给高精度时序建模带来挑战[1。（剩余13452字）