基于地基云图与气象数据的辐照度轻量化预测

A Lightweight Solar Irradiance Prediction Model Based on Ground-based Cloud Images and Meteorological Data

针对总水平辐照度(Global Horizontal Irradiation,GHI)的短时快速预测问题,提出了一种基于深度学习技术的全天空太阳总辐照度轻量化预测模型.首先,区别于传统的人为处理云图区分特征方法,采用卷积神经网络结构MobileNet自动提取地基云图特征.其次,通过对9项气象参数进行特征清洗,权重连接,建立了短时辐照度的预测方案.最后,将上述模型迁移至微型终端上,在神经网络计算棒的辅助下,实现了GHI的轻量化快速预测.结果表明,相较于传统使用地基云图的时延神经网络预测模型,此轻量化GHI预测模型标准均方根误差降低了5.5%,较未轻量化模型预测速度提升了71%,可供家庭个体光伏用户进行快速预测.

Aiming at the short-term and rapid prediction of Global Horizontal Irradiation(GHI),a lightweight prediction model of total sky solar irradiance based on deep learning technology is proposed.First of all,different from the traditional artificial processing of cloud image distinguishing features,it uses the convolutional neural network structure MobileNet to automatically extract ground-based cloud image features.Secondly,a short-term irradiance prediction scheme was established by performing feature cleaning on 9 meteorological parameters and connecting weights.Finally,the above model was migrated into a Raspberry Pi system,and with the aid of the neural network computing stick,the lightweight and rapid prediction of GHI was realized.The results show that compared with the traditional neural network prediction model using ground-based cloud images,the standard root mean square error of this lightweight GHI prediction model is reduced by 5.5%,and the prediction speed is increased by 71%.It can be used for rapid prediction of individual household photovoltaic users.