基于RNN的短期太阳辐照度预测算法研究  

Research on Short Term Solar Radiation Prediction Algorithm Based on RNN

作　　者：马景奕[1] 王帅[2] 闫文君 李雅文 田瑜 Ma Jingyi;Wang Shuai;Yan Wenjun;Li Yawen;Tian Yu(China Meteorological Administration Meteorological Cadre Training Institute Gansu Branch,Lanzhou 730020,China;National Meteorological Information Centre,Beijing 100081,China)

机构地区：[1]中国气象局气象干部培训学院甘肃分院,兰州730020 [2]国家气象信息中心,北京100086

出　　处：《科技通报》2022年第5期16-22,共7页Bulletin of Science and Technology

基　　金：国家气象信息中心信息网络安全与“信创”技术研发创新团队(NMIC-202011-05)攻关任务资助。

摘　　要：预测太阳辐照度对于有效及时的利用可再生能源至关重要。本文旨在研究递归神经网络(RNN)的5个变体,并得出有效可靠的5 min短期太阳辐照度预测模型。5个RNN网络分别是长期短期记忆(LSTM),门控循环单元(GRU),简单RNN,双向LSTM(Bi-LSTM)和双向GRU(Bi-GRU);前3个类别是单向的,后2个类别是双向的RNN模型。基于24个月连续采集的相关天气与辐照数据对5个网络模型进行训练与测试,研究不同参数以及模型结构下预测精度与误差的变化,最终得到最优的模型种类与结构。实验表明模型的深层次的体系结构会产生显著效果,同时,与单向预测相比,Bi-LSTM和Bi-GRU能提供更准确的预测。Bi-GRU模型提供了最低的RMSE和最高的R^(2)值,分别为46.1和0.958;此外,双向RNN显示出较高的鲁棒性与非线性表达能力。It is very important to predict the amount of solar radiation for the effective and timely use of renewable energy. The purpose of this paper is to study five variants of recurrent neural network(RNN) and obtain an effective and reliable 5-minute short-term solar irradiance prediction model. The five RNN networks are long-term short-term memory(LSTM), gated cycle unit(GRU), simple RNN, bi-directional LSTM(BI LSTM) and bi-directional GRU(BI GRU);the first three categories are unidirectional, and the last two categories are bidirectional RNN models. Based on 26 months of continuous collection of relevant weather and radiation data, five network models were trained and tested to study the changes of prediction accuracy and error under different parameters and model structure, and finally the optimal model type and structure were obtained. Experiments show that the deep-seated architecture of the model can produce significant effect, and Bi LSTM and Bi GRU can provide more accurate prediction than one-way prediction. Bi GRU model provides the lowest RMSE and the highest R^(2), which are 46.1 and 0.958 respectively;in addition, bi-directional RNN and simple RNN model show high robustness and nonlinear expression ability.

关 键 词：太阳辐照 递归神经网络 深度学习 

分 类 号：P422.1[天文地球—大气科学及气象学]