基于特征时间双注意力机制的短期光伏发电预测深度学习模型研究  

A Deep Learning Model for Short-term Photovoltaic Power Generation Forecasting Based on Dual Feature-Temporal Attention Mechanism

作　　者：王建军[1] 潘佳音 赵珍珠 肇启迪 WANG Jianjun;PAN Jiayin;ZHAO Zhenzhu;ZHAO Qidi(School of Economics and Management,North China Electric Power University,Beijing 102206,China)

机构地区：[1]华北电力大学经济与管理学院,北京102206

出　　处：《智慧电力》2025年第4期81-87,共7页Smart Power

基　　金：国家自然科学基金资助项目(72071077)。

摘　　要：针对光伏发电预测中存在的输入特征变量选择不精准和长时间历史信息难以捕捉等问题,提出一种基于特征时间双注意力机制的短期光伏发电功率预测深度学习模型(DA-GRU)。利用特征注意力机制挖掘不同影响特征对光伏发电功率的重要程度,通过时间注意力机制衡量历史信息在不同时间点上的重要性,从而有效捕捉长时间序列上的变化趋势。算例分析表明,所提模型在各项评价指标上均优于其它对比模型,说明其对复杂非线性光伏发电功率数据有较好的适应性。To address challenges in photovoltaic(PV)power forecasting,such as imprecise selection of input feature variables and difficulty in capturing long-term historical information,this paper proposes a deep learning model named dual attention gated recurrent unit(DA-GRU)for short-term PV power prediction.The model employs a feature attention mechanism to evaluate the significance of different influencing features on PV power generation and a temporal attention mechanism to weigh the importance of historical information at different time points,thereby effectively capturing long-term temporal trends.Case study analysis demonstrates that the proposed model outperforms other comparative models across all evaluation metrics,highlighting its superior adaptability to complex nonlinear PV power data.

关 键 词：光伏发电预测 双阶段注意力机制 门控循环单元 

分 类 号：TM615[电气工程—电力系统及自动化]