\r董 娜，常建芳，吴爱国\r
\r
AuthorsHTML:\r董 娜，常建芳，吴爱国\r
\r
AuthorsListE:\rDong Na，Chang Jianfang，Wu Aiguo\r
\r
AuthorsHTMLE:\rDong Na，Chang Jianfang，Wu Aiguo\r
\r
Unit:\r天津大学电气自动化与信息工程学院，天津 300072\r
\r
Unit_EngLish:\rSchool of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China\r
\r
Abstract_Chinese:\r由于不同气象条件会影响太阳辐照度的有效利用，这制约了太阳能的应用和发展．为了基于不同站点不同采样时刻的气象属性预测中尺度站的太阳能辐照度，依据传统卷积神经网络的框架，建立了一种新型的卷积神经网络结构并用于太阳能辐照度预测．为了缓解新型网络由超参数选取不当导致预测性能差的问题，利用融合算法对新型网络的超参数进行优化．为了提高融合优化算法的全局搜索能力，引入帐篷映射对粒子的初始位置和初始速度进行混沌初始化．首先，导入训练集更新新型卷积神经网络框架，训练结束后导入验证集检验当前模型参数下新型卷积框架的性能．其次，混沌融合算法依据新型卷积神经框架在验证集上的预测性能更新模型的超参数．对更新模型的超参数多次检验，直至最优的预测模型在验证集上的性能趋于收敛．最后，输出模型的最优超参数，建立太阳能辐照度预测模型．基于气象实测数据建立太阳能辐照度预测实验，引入其他两种预测方法进行对比仿真研究，并尽可能复现了Eustaquio and Titericz 团队的预测方法(GBRT)作为太阳能辐照度预测性能的评估基准．实验数据表明：混沌融合算法可以有效地提高新型卷积神经网络的预测性能，所提出预测方法的全年太阳能辐照度的均方误差较GBRT 降低25.9%，绝对平均误差较GBRT 降低了10.7%；全年太阳能辐照度平均误差率降低了18.4%，误差率小于
0.1 的样本量增加了21.1%．\r
\r
Abstract_English:\rSeasonal factors，such as season，climate，cloud density，and other climatic factors，restrict the stability and applications of solar irradiation. In this paper，a novel prediction method based on the framework of traditional convolutional neural networks is proposed to predict the solar irradiance of mesoscale stations based on meteorological data obtained at different sampling moments from different global ensemble forecast system(GEFS) stations. To alleviate the issue of over-fitting or under-fitting caused by improper selection of hyper parameters，the chaotic hybrid algorithm is utilized to optimize the hyper parameters of the novel framework. In addition，a tent map is utilized to improve the global search ability of the hybrid optimization algorithm. First，a training set is constructed to update the novel convolutional neural network framework，after which a validation set is imported to test the performance of this framework under the current hyper parameters. Next，the chaotic hybrid algorithm updates the hyper parameters according to the prediction performance of the novel framework on the verification set until the performance of the optimal prediction model on the verification set tends to converge. Finally，the optimal hyper parameters are utilized to develop a solar irradiance prediction model. A solar irradiance prediction experiment is established based on meteorological data，and two machine learning prediction methods，as well as the prediction method (GBRT)of Eustarquio and Titericz，are introduced to enable comparison of simulation results. The experimental data demonstrate that the chaotic hybrid algorithm can effectively improve the prediction performance of the novel framework；specifically，the mean squared error(MSE) of the proposed method is 25.9% lower than that of GBRT，the mean absolute error (MAE) of the former is reduced by 10.7% compared with that of the latter，and compared with GBRT，the average error rate of the proposed method is reduced by 18.4%. Samples with error rates less than 0.1 of GBRT account for 58.525%，while that of the proposed method account for 70.89%，which increased by 21.1%.The experimental results verify the accuracy and effectiveness of the proposed prediction method in solar irradiance prediction.\r
\r
Keyword_Chinese:卷积神经网络；混沌融合算法；参数优化；太阳能辐照度预测\r

Keywords_English:convolution neural network；chaotic hybrid algorithm；parameter optimization；solar irradiance prediction\r


PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6316