基于小波分解的时序预测模型mWDLNet及其应用研究

Research on Application of Time Series Forecast Model mWDLNet Based on Wavelet Decomposition

各种类型的深度神经网络模型已被应用到时序分析中,但基于频域的神经网络与时域的线性模型融合仍然缺乏有效的模型.提出一种基于多级小波分解的深度网络和差分自回归移动平均模型相融合的方法(mWDLNet),时序信号经小波分解到频域,由卷积神经网络和长短期记忆网络提取时序信号的空间和时间维度特征,同时利用差分自回归移动平均模型(ARIMA)解决神经网络模型的尺度不敏感问题,最后融合两部分的输出结果,提高了预测的准确性.并通过实验验证了mWDLNet模型添加小波分解、卷积模块以及融合线性预测的有效性.将提出的模型应用于北京气象数据集,进行PM2.5浓度预测,并与常用的时序预测模型进行对比分析,结果表明,提出的mWDLNet模型能达到更好的预测结果.

Various types of deep neural network models have been used in time series analysis.But the fusion of neural networks based on the frequency domain and linear models still lacks effective models.A method(mWDLNet)based on multi-level wavelet decomposition of deep network and ARIMA is proposed.The original time series is decomposed into sub-sequences of different frequencies.One-dimensional convolutional neural network(CNN)extracts the spatial features,LSTM extracts long and short-term temporal features,and ARIMA solves the scale insensitivity of the neural network model.Finally,the final prediction is the weighted sum of the output of the neural network and the linear model.And experiments verify the effectiveness of adding wavelet decomposition,convolution module and fusion linear prediction to mWDLNet model.The proposed model is used to predict PM2.5 concentration in Beijing,and compare and analyze the prediction results with the common time series prediction models.The mWDLNet model can yield better prediction results.