A Deep Learning Framework for Non-stationary Time Series Prediction

A Deep Learning Framework for Non-stationary Time Series Prediction

연구 분야: Artificial Intelligence

학회: 2022 3rd International Conference on Computer Vision, Image and Deep Learning & International Conference on Computer Engineering and Applications (CVIDL & ICCEA)

In non-stationary time series, there are data bursts, which brings challenges to accurately predict data. This paper proposes a deep learning framework for non-stationary time series prediction. In this framework, the first-order and second-order difference and decomposition of the original time series are first made respectively, so as to generate five new time series, which are as the input of the framework. Then, a prediction model is constructed sequentially by GRU (gated recurrent unit) and FCN (fully-connected network) networks to predict and fit data. Finally, a two-stage training mode is designed, which first predicts the trend and component, and then fits with the cycle to produce the prediction data. The experiments are tested on the public air quality dataset, and the results show that our approach can accurately predict the non-stationary time series, especially overcomes the lag, and achieves better performance than typical statistics methods and deep learning models.