DSENet: A Novel Dual-Stream Enhancement Network for Multi-Scale Non-Stationary Time Series Forecasting

Accurately capturing local variations in long series has always been one of the most challenging problems in time-series forecasting especially in medical signals, where local variations often indicate pathological events. Our study reveals a previously overlooked key bottleneck in this field: traditional global and local branches learn similar representations, leading to strong feature coupling and reduced sensitivity to local variations. To address this challenge, we propose the novel Dual-Stream Enhancement Mechanism, which structurally enlarges the difference between global and local patterns, enabling weak interactions between the two. Based on this idea, we introduce a new baseline model for blood glucose prediction: Dual-Stream Enhancement Network (DSENet), which fundamentally alleviates the problem of excessively strong coupling between global and local features. Experimental results show that our model achieves SOTA performance on multiple public datasets. Moreover, benefiting from extremely low computational cost, our model demonstrates strong application potential and can serve as a baseline model in multiple domains in the future.

Many important signals, such as blood glucose measurements, change over time through both slow trends and short local fluctuations. Predicting these signals is difficult because small local changes can carry important information, especially in medical settings where they may indicate abnormal or risky physiological events. Existing forecasting models often try to learn long-term patterns and local details at the same time, but these two parts can become too similar inside the model. When this happens, the model may become less sensitive to subtle local variations that matter for accurate prediction. We address this problem with a Dual-Stream Enhancement Mechanism, which encourages the model to learn long-term trends and local changes in more distinct ways while still allowing them to interact. Based on this mechanism, we build DSENet, a forecasting model for blood glucose prediction. Experiments on multiple public datasets show that DSENet achieves better prediction accuracy than existing baseline methods. Because the model also has low computational cost, it may be useful for practical glucose forecasting systems and may serve as a simple baseline for other long time-series forecasting tasks.