摘要
针对双阶段注意力自编码神经网络(DA-RNN)时间序列预测算法对随机数据预测效果较差和长时间预测中存在的累积误差问题进行改进,设计了一种基于生成式对抗网络(GAN)的时间序列预测算法。该算法以DA-RNN网络为生成网络,利用生成网络和判别网络之间的互补特性,消除DA-RNN网络对于长时间预测过程中存在的累积误差问题;引入多维注意力模型改进DA-RNN网络,并利用稀疏映射函数改进多维注意力模型;改进网络优化目标,通过探索目标序列在不同分位数下分布的形式,提升网络对于随机数据的预测精度。通过在公开数据集上测试,对算法的的准确性和有效性进行验证,结果表明:本文算法与DA-RNN算法相比,累积误差有明显降低,且对于随机数据的预测精度有显著提高。
In order to improve random data prediction performance and reduce accumulated error in long-term prediction of the dual-stage attention-based recurrent neural network(DA-RNN), a time series prediction algorithm based on generative adversarial networks(GAN) is designed in this paper. The algorithm uses the DA-RNN network as the generative network, and uses the complementary characteristics between the generative network and the discriminant network to eliminate the cumulative error of the DA-RNN network in the long-term prediction process;it introduces a multi-dimensional attention model to improve the DA-RNN network, and uses the sparse mapping function to improve the multi-dimensional attention model. In addition, it improves the network optimization target and enhances the network’s prediction accuracy for random data by exploring distribution of the target sequence in different quantiles.Through testing on public data sets, the accuracy and effectiveness of the algorithm are verified. The results show that compared with the DA-RNN algorithm, the cumulative error of the algorithm is significantly reduced, and the prediction accuracy of random data is significantly improved.
作者
闫保中
苏邓军
YAN Baozhong;SU Dengjun(College of Intelligent Science and Engineering,Harbin Engineering University,Harbin 150001,China)
出处
《应用科技》
CAS
2022年第2期114-118,126,共6页
Applied Science and Technology
