Deep Learning Time Series Prediction Model Based On Self-decomposition And Self-attention

Time series forecasting plays a key role in numerous fields such as economy,finance,energy,transportation,medical treatment,meteorology and commerce.It empowers people to foresee opportunities and serves as guidance for decision-making.In the field of time series forecasting,multi-variable and multi-step forecasting forms one of the most challenging tasks.Errors may accumulate as the forecast step increases.Some auxiliary variables are usually needed to aid the forecast of a target variable.How to characterize the relationship between the target variable and auxiliary variables is an important research topic.In recent years,with the emergence and remarkable achievements of deep learning in many fields,in the literature many attempts have also been made to its application in time series forecasting.Compared with traditional methods,deep neural network-based methods can better capture highdimensional features of time series data.However,due to the complex modalities inherent in real world time series,feature engineering is required to obtain good prediction results.Traditional manual feature engineering lacks the versatility and robustness hence usually one time series calls for its specific forecasting model and feature engineering.Moreover,current deep learning based methods usually suffer from interpretability issue,so the “black box” results lead to a lack of confidence.To deal with the above issues forms the motivation of the thesis.In the model we propose a self-decomposition mechanism and a self-attention mechanism,where the self-decomposition network is used to automatically perform feature decomposition and feature extraction,and the self-attention module is to capture varying importance of the data.In addition,the thesis also proposes a variable selection network for adaptively treating the features: suppressing the unimportant features whereas enhancing the essential ones.Our model is based on automatic feature engineering and has strong feature expression capability.It can look into the “black box” of the deep neural network by feature visualization of the self-decomposing network and the variable selection network.Moreover,our model is general in its demonstrated capability of dealing with many different types of time series data.Its forecasting accuracy also over performs typical DNN methods in the literature.