Long Sequence Time Series Forecasting Using Spectral ConvMixer Alongside Weak-stationarizing and Non-stationarity Restoring Blocks
Supervised by: Prof. Sang-Woong Lee
Abstract: Real-life time series data accommodate some intricacies that hamper the possibility of precise forecasts. While adhering a non-stationary nature, real-life time series data also exhibit complex patterns in both of inter and intra-series relationships. These innate features of real-life time series data are hindrance to the task of forecasting. Many time series forecasting models do not regard these hurdles while making forecasts and are likely to endure erroneous results. This thesis proposes a deep learning architecture for time series forecasting which addresses the non-stationarity of time series while modeling the complex inter-series and intra-series patterns of time series data. To deal with the non-stationarity of time series this thesis proposes methods called ’Weak-stationarizing’ and ’Non-stationarity Restoring’ to deal with the problem of distribution shift. These methods remove and restore the non-stationary components of individual data points respectively, as needed. Then to model the inter series and intra series relationship of real-life time series data, the architecture obtains the spectral decompositions of the weak-stationary time series to learn features for making forecasts. The architecture utilizes a series of mixer layers to learn features from the obtained spectral decomposition of weak-stationary time series. To demonstrate the potency of the suggested model, it has been compared with existing state-of-the-art models on six real-world datasets, which address 5 diverse fields of application: energy, economics, traffic, weather, and health. In each dataset, the suggested model performs better or has comparable results to that of the existing state-of-the-art models.
Thesis      Slides

Under Review

Select Publications