基于深度学习的FY-4A图像地表太阳辐照度预测  

Prediction of Surface Solar Irradiance in FY-4A Images Based on Deep Learning

作　　者：姚蕊 刘小芳 张杰 郭旭萍 YAO Rui;LIU Xiaofang;ZHANG Jie;GUO Xuping(School of Computer Science and Engineering,Sichuan University of Science&Engineering,Yibin 644000,China;Information Technology Department,Shanxi Yuncheng Vocational and Technical College of Agriculture,Yuncheng 044000,China)

机构地区：[1]四川轻化工大学计算机科学与工程学院,四川宜宾644000 [2]山西运城农业职业技术学院信息技术系,山西运城044000

出　　处：《四川轻化工大学学报（自然科学版）》2024年第4期58-67,共10页Journal of Sichuan University of Science & Engineering(Natural Science Edition)

基　　金：教育部高等教育司产学合作协同育人项目(202101038016);高层次创新人才培养专项资助项目(B12402005);四川轻化工大学人才引进项目(2021RC16)。

摘　　要：针对现有地表太阳辐照度预测方法存在的建模复杂、范围有限、精度低、四季差异大以及多任务协同预测困难等问题,提出了一种基于Encoder-Decoder框架融合卷积神经网络、自注意力机制和双向门控循环单元的深度学习模型,用于地表太阳辐照度预测。首先,利用分块技术处理图像,在编码器中构建卷积层并增大输出特征向量维度以提取图像序列的空间特征,显著扩展预测范围,强化特征提取能力;然后,采用自注意力机制来编码输入序列,结合双向门控循环单元从全局角度捕捉序列语义上下文时序信息,减小局部感受野的限制;最后,采用均方误差和结构相似性加权衡量损失,强化地表太阳辐照度中短时预测效果。结果表明,该方法在简化模型输入参数的同时,均方误差降低至21.75(W/m^(2))^(2),结构相似性高达88.46%,相较于传统算法有较大提升,能够有效提高地表太阳辐照度预测性能。To address the issues such as complex modeling,limited prediction range,low accuracy,significant seasonal variations,and difficulties in multi-task collaborative prediction in existing methods for surface solar irradiance prediction,a deep learning model based on the Encoder-Decoder framework combined with convolutional neural networks,self-attention mechanism and bidirectional gated recurrent units has been proposed,which is used to predict the surface solar irradiance.Firstly,the images are processed using the block partitioning technology,the convolutional layers are constructed in the encoder and the output feature vector dimension is increased to extract spatial features of image sequences,significantly expanding the prediction range and enhancing feature extraction capabilities.Subsequently,the self-attention mechanism is employed to encode the input sequence and the global contextual temporal information is captured by the bidirectional gated recurrent units to reduce the limitations of local receptive fields.Finally,to improve short-term prediction accuracy of surface solar irradiance,a weighted loss function based on mean squared error and structural similarity is adopted.The results suggest that the inputted parameters are simplified and the mean squared error is remarkably reduced to 21.75(W/m^(2))^(2) with a high structural similarity of 88.46%.Compared to traditional algorithms,the proposed method exhibits obvious improvements,which can effectively enhance the prediction performances of surface solar irradiance.

关 键 词：地表太阳辐照度预测 卷积神经网络 自注意力机制 双向门控循环单元 深度学习 图像序列 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] P422.12[自动化与计算机技术—计算机科学与技术]