Analysis And Prediction Of Complex Time Series Based On Deep Learning

Time series analysis is an important method of revealing natural phenomena and understanding the objective world.In recent years,time series prediction has become a popular research area and is widely used in various fields such as financial data analysis,air quality assessment,and traffic flow prediction.With the continuous improvement of science and technology,people have transformed from univariate time series analysis to multivariate complex time series analysis,and tried to analyze multiple components of the system to improve the prediction accuracy,but the existing complex time series prediction methods still have the following challenges: First,the interference of abnormal data,most prediction models have very low tolerance for abnormal data,and the imbalance and non-labeling of complex time series make abnormal data difficult detected;Second,the interference of noise components,the noise components cause the self-similarity of the time series to be destroyed,and the blurring of the boundaries between the useful components and the noise components in the complex time series leads to the risk of losing useful components when de-noising;Third,the complex interdependence between multiple components of complex time series makes prediction difficult.Aiming at these problems,this thesis conducts research from three aspects,which are complex time series anomaly detection,noise reduction and prediction.The specific work and research results are as follows:(1)Aiming at challenge 1,this thesis proposes an anomaly detection method based on prelabeling and multi-objective generation adversarial network.The pre-labeling method is used to label normal data and the multi-objective generate network is trained to generate abnormal data to solve the problem of imbalance and non-labeling of complex time series.Finally,the anomaly detection classifier is obtained through the confrontation game between the generative network and the discriminat network.The effectiveness of the proposed anomaly detection method is proved on the public datasets.(2)Aiming at challenge 2,this thesis proposes a noise reduction method based on complete ensemble empirical mode decomposition with adaptive noise and adaptive threshold.Distinguish the useful components and noise components in the frequency domain through time series decomposition,and construct an adaptive heuristic threshold function to keep the useful components as much as possible while reducing noise.The effectiveness of the proposed noise reduction method is proved on the simulated time series with different noise intensities.(3)Aiming at challenge 3,this thesis proposes a complex time series ensemble prediction method based on attention stacked long short-term memory neural network,and takes anomaly detection and noise reduction as a preprocessing part.Extracting complex time series features by stacked long and short-term memory neural networks,introducing attention mechanism to obtain weighted feature inputs,and enhancing prediction ability through ensemble algorithm.Taking the air quality index PM2.5 of Beijing and Guangzhou as the analysis objects,the effectiveness of the proposed prediction model is proved,and the effectiveness of the abovementioned anomaly detection method and noise reduction method is proved too.