<p>Time series forecasting plays a significant role in fields such as traffic planning, weather prediction, economics, and energy management. However, time series in real-world scenarios typically exhibits complex multi-scale dynamic characteristics: coarse-grained sequences can depict long-term trends but lack local details, while fine-grained sequences can capture short-term fluctuations yet are vulnerable to noise interference. How to effectively integrate long-term dependencies and short-term changes in the forecasting process remains a key challenge. To address this, this paper proposes ScaleMixNet, a deep learning framework for multi-scale modeling. This framework consists of two core modules. The first is the multi-scale adaptive synergistic filtering block (MASFB), which is used to extract and fuse frequency components across scales. The second is the adaptive convolutional filtering block (ACFB), which is designed to dynamically integrate local temporal fluctuations at multiple scales, thereby improving forecasting accuracy. In addition, we design an adaptive hybrid loss function that overcomes the limitations of a single loss function in the optimization process by dynamically balancing the weights of MSE and MAE, enhancing the robustness and generalization ability of the model. Experimental results on multiple real-world datasets spanning transportation and energy domains demonstrate that ScaleMixNet outperforms existing methods like TimeMixer and iTransformer in prediction performance, with an average relative reduction of 3.4% in MAE, verifying its effectiveness and superiority in complex time series modeling.</p>