Abstract: Time series forecasting plays a crucial role in various real-world applications such as energy management,traffic flow forecasting,and meteorological analysis.However,the presence of distribution shift and long-term dependency in time series data continues to limit the performance of both traditional methods and existing deep learning models in long-range forecasting.To address these challenges,this paper proposes an innovative model named D-LINet.The proposed model integrates the distribution normalization capability of the Dish-TS framework with the efficiency of linear mappings.By employing dual-direction normalization and dual-linear-layer designs,it effectively mitigates distribution shifts in both input and output spaces,while significantly enhancing the capture of periodic and trend-related features.A comprehensive evaluation of D-LINet on multiple real-world datasets demonstrates that,for both short- and long-term forecasting,D-LINet consistently achieves lower MSE and MAE compared to mainstream models such as Transformer,Informer,Autoformer and DLinear.In addition,experiments investigate the influence of input window length and the incorporation of prior knowledge on forecasting performance,providing valuable insights for subsequent model optimization.Overall,this study offers a novel solution to address complex distribution shifts,contributing to improved accuracy and robustness in time series forecasting.

Key words: Time series forecasting(TSF), Distribution shift, Dual normalization, Linear mapping, Periodicity and trend modeling