基于改进VMD和注意力机制的混沌时间序列预测

Prediction of Chaotic Time Series Based on Adaptive V ariational Modal Decomposition and Self attention Mechanism

采用神经网络预测混沌时间序列时,已有研究忽略了混沌时间序列的物理信息混叠现象,难以满足预测时混沌时间序列特征提取的需求,而变分模态分解的结果受参数影响较大,合适的参数能够大大提升变分模态分解的效果,故提出一种模拟退火算法优化的自适应变分模态分解算法,解决参数选择问题。长短期记忆神经网络适用于时序信息的处理,但易忽略输入之间的关联信息,而自注意力机制可加强数据内部关联,有利于重要特征的提取,因此将自注意力机制融入长短期记忆网络。通过引入这两种方法形成融合神经网络模型,采用真实混沌时间序列验证了所提出方法的有效性。对比实验结果表明:所提出的融合神经网络模型可有效提取混沌时间序列中隐含的动力学特征,显著改进神经网络的注意力配比,有效解决预测中出现的超前现象,从而大幅提高了混沌时间序列预测的精确性和稳定性。

When neural networks are used to predict chaotic time series,existing studies have ignored the physical informat ion aliasing phenomenon of chaotic time series,so it is difficult to meet the demand for feature extraction of chaotic time series during prediction,and the resuls of variational mode decomposi-tion are greatly affected by parameters,as appropriate parameters can greatly improve the effect of varia-tional mode decomposition,an adaptive variational mode decomposition algorithm optimized by simulated annealing algorithm is proposed to solve the problem of parameters selection.The long-short+term memory neural network is suitable for the processing of temporal information,but it is likely to ignore the correla-tion information between inputs,and the self attention mechanism can strengthen the interal correlation of data,which is conducive to the extraction of important features,thus the self attention mechanism is inte-grated into the long-term short-term memory network.The fusion neural network model is formned by intro-ducing these two methods,and the effectiveness of the proposed method is verified by real chaotic time se-ries.The comparative experiments show that the proposed fusion neural network model can effectively ex-tract the hidden dynamic features in the chaotic time series,improve the attention ratio of the neural network signifcanly,and effectively solve the advance phenomenon in prediction,so as to greatly improve the accuracy and stability of chaotic time series predication.