Based on hypernetworks and multifractals: Deep distribution feature fusion for multidimensional nonstationary time series prediction

It is commonly assumed that time series data follows a stationary distribution in time series prediction tasks, while the reality is that distribution drift issues are widespread. This study addresses the prediction of multi-dimensional non-stationary time series by proposing a solution that incorporates feature fusion based on classical fractal theory, specifically the multifractal spectrum width, and a hyper-network model. The HFMF (Hypernetwork Feature-Fusion Multidimensional Forecasting) network is designed, and innovative deep convolution modules, Mdff (Multidimensional deep feature fusion) and Dmff (Deep multifractal feature fusion), are introduced to fuse distribution information and time series information at the hyper-layer of the hyper-network. In the experimental section, extensive experiments are conducted on five publicly available datasets, including model performance testing, ablation experiments, generalization experiments, and other experiments. The results confirm the superior performance of our method in predicting multivariate non-stationary time series, surpassing other state-of-the-art techniques.