基于生成对抗网络的太阳辐照度预测模型

Solar Irradiance Prediction Model Based on Generation Adversarial Networks

为了进一步提高辐照度预测的精度,提出一种改进的太阳辐照度预测深度混合模型,即自编码器条件深度卷积生成对抗网络(auto encoder-conditional-depth convolution generation adversarial networks,AE-C-DCGAN)模型。该混合深度模型集成了自编码器(auto-encoder,AE)和条件深层卷积生成对抗网络(conditional-depth convolution generation adversarial networks,C-DCGAN)模型,AE用于减少辐照度序列的非线性和非平稳性,C-DCGAN由卷积和反卷积层堆叠而成,用于时间特征提取和预测工作。与7种较流行模型的实验对比结果表明,AE-C-DCGAN模型的误差最小;与次优预测模型相比,AE-C-DCGAN模型的四季预测结果的平均绝对误差、平均绝对百分比误差和均方根误差分别平均降低了4.84%、8.49%、7.32%,表明本文提出的模型能够有效提升太阳辐照度的预测精度。

This paper proposes an improved hybrid depth model,auto encoder-conditional-depth convolution generation adversarial networks(AE-C-DCGAN),to improve the accuracy of irradiance prediction.It integrates auto-encoder(AE)and conditional-depth convolution generation counter adversarial network(C-DCGAN)model.AE is used to reduce the nonlinearity and nonstationarity of the irradiance series,and C-DCGAN is composed of convolution and deconvolution layers for time feature extraction and prediction.Compared with seven popular models,the error of this model is the smallest;compared with the suboptimal prediction model,the MAE,MAPE and RMSE of the four seasons forecast results of this model is 4.84%,8.49%and 7.32%lower respectively.All this indicates that the model proposed in this paper can effectively improve the prediction accuracy of solar irradiance.