| Time series forecasting has been widely used in various fields,such as engineering and finance,especially Long Sequence Time Series Forecasting(LSTF)that the length of the forecast time series is relatively large,is especially required to accurately capture the long-term dependencies of the input and output time series.It is very meaningful for long-term planning and early warning,such as long-term planning of power and energy consumption.However,it is difficult for existing models to capture reliable long-term dependencies in their complex patterns for learning when making Long Sequence Time Series Forecasting for non-stationary time series with coexisting trend and periodic characteristics,and Long Sequence Time Series Forecasting also has corresponding requirements for the time complexity and memory consumption of the model running,which is prone to problems like unpredictable memory overflow.Aiming at the above problems,this thesis proposes a Long Sequence Time Series Forecasting Model Based on Time Series Decomposition and Probabilistic Sparse SelfAttention.The main work is as follows:(1)A Time Series Decomposition Model named Robust SS that integrates singular spectrum analysis is proposed.For the conventional Long Sequence Time Series Forecasting Model,it is difficult to capture long-range information in the complex mode of non-stationary time series.The time series decomposition model is utilized to decompose the trend and periodic characteristics,reduce the non-stationarity of the time series,and help the subsequent Long Sequence Time Series Forecasting Model to better capture related information.First,the method of bilateral filtering is employed to remove the noise,and the abnormal data is removed on the premise of retaining the trend mutation information,and then the minimum absolute deviation is leveraged to solve the corresponding regression problem to extract the trend feature.After the trend feature is extracted,the non-local seasonal filtering method is applied to extract the periodic feature,and the remainder is obtained by iterative update.Finally,using Singular Spectrum Analysis method to analyze the main periodic components of the periodic subsequence,and reconstruct the main periodic components to obtain a new periodic subsequence.Multiple sets of experimental results show that the Robust SS model proposed in this thesis has an average improvement of 3.051% in reducing nonstationarity and an average reduction in decomposition error of 1.950% when compared with the baseline model Robust STL.(2)A Long Sequence Time Series Forecasting Model named SS-Informer with Probabilistic Sparse Self-Attention is proposed.The attention-based long-term prediction model can learn how to capture long-range information from non-stationary time series,but it has high time complexity and high memory consumption.Using the non-stationary time series decomposed by the Robust SS model as input,the Transformer long-term prediction model is improved,and the query vector of the attention mechanism is measured for sparsity.Then,the dynamic decoder of step-bystep prediction is replaced by a generative decoder to complete the prediction of the required length for long-term prediction in one step.Finally,the results are reconstructed for the three groups of subsequences to reduce the overall runtime complexity.Multiple sets of experimental results show that the SS-Informer model proposed in this thesis reduces the univariate prediction error by 10.087% on average,and the multivariate prediction error by 2.877% on average when compared with the baseline model Informer.It is verified that the SS-Informer model successfully reduces the error of long-term prediction of non-stationary time series.(3)A Long Sequence Time Series Forecasting Model named SSL-Informer fused with residual learning is proposed.The residual sequence is learned,and the hidden sequence information that is missed when SS-Informer makes LSTF is mined to further reduce the long-term prediction error of SS-Informer.Use the original input time series and the fitted input time series of SS-Informer to generate residual sequence,and then perform pure random sequence test on the residual sequence,and reconstruct the forecasting result after LSTM residual learning and SS-Informer forecasting result to obtain the final forecast result.Multiple sets of experimental results show that the Long Sequence Time Series Forecasting error of the SSL-Informer model is reduced by an average of 0.715% in univariate forecasting errors and 1.250% in multivariate prediction errors when compared with the SS-Informer model.It is verified that the SSL-Informer model has better long-term prediction ability than SS-Informer. |
