Deep learning models have achieved substantial advancements in time series forecasting, leveraging their impressive capability to capture sequential dependencies. However, time series data often exhibits non-stationarity, where the data distribution undergoes rapid changes over time, posing significant challenges for accurate prediction. Some studies have proposed normalizing data within subsequences to eliminate non-stationarity and adopting a statistical forecasting module to model the evolution trend of the original time series’ statistical characteristics. Based on the predicted future statistical characteristics, denormalization is performed to restore stationarity, achieving accurate predictions of future distributions. This statistical forecasting approach has limitations in handling outliers in statistical features, resulting in insufficient model robustness. To address this issue, the Slice Adaptive Normalization module with Huber loss (H-SAN) is proposed, which effectively enhances the model’s robustness against outliers and improves prediction performance. Additionally, while this method employs local normalization, it utilizes long-sequence modeling for feature extraction, neglecting the inter-slice dependencies. Furthermore, this paper proposes a dual-token input strategy, which achieves simultaneous extraction and effective fusion of both global and local features based on the dual scales of input tokens. A network for non-stationary time series forecasting based on Distribution Shifts and Data Dependency Network (DSDDNet) is constructed in this paper. Experimental results demonstrate that DSDDNet significantly outperforms existing benchmark models on standard datasets, validating its effectiveness and superiority.