生成对抗网络在风电功率预测中的应用

Application of generation countermeasure network in wind power forecasting

针对极端气象条件所产生的异常数据使得预测结果在极端气象情况下存在预测精度低和异常识别错误等问题,提出一种基于对抗生成式网络的风电功率预测方法.该方法基于联合分布KL散度,分析并明确正常数据与异常数据的联合分布,经过不同模式下的对抗训练,生成风电功率数据集,实现多种气象条件下风电功率预测.研究结果表明:本文所采用的对抗生成网络采用了ReLU激活函数,其预测误差具有更好的收敛性.结论证明多任务学习可适用于多种气象条件下的风电功率预测,并且在减少数据清洗的同时,提高风电功率的预测精度.

Aiming at the abnormal data generated by extreme weather conditions,the prediction results have problems such as low prediction accuracy and abnormal recognition errors in extreme weather conditions,this paper proposes a wind power prediction method based on a confrontation generative network.The method is first based on the joint distribution KL Divergence,the joint distribution of normal data and abnormal data is analyzed and clarified,and then wind power data sets after confrontation training in different modes is generated to achieve wind power prediction under various weather conditions.The research results show that the confrontation generation network uses ReLU activation function,and its prediction error has better convergence.The conclusion proves that multi-task learning can be applied to wind power prediction under a variety of weather conditions,and it can improve the prediction accuracy of wind power while reducing data cleaning.