DMNet: time series forecasting based on dynamic attention and multi-branch structure

Time series forecasting is applicable in wide range of fields, for example meteorology, traffic monitoring, investment, energy production, and consumption. Over the past few decades, both statistical and machine learning approaches were utilized to resolve time series forecasting challenges. In contrast, these methods have problems including lack interpretability or exhibit less than satisfactory performance. Therefore we propose a DMNet model based on the FDNet model. First, the model considers the introduction of a dynamic attention module, which aims to enhance the effectiveness of the model by focusing on global information to compensate for the shortcomings of capturing only local information. Second, the model adds a multi-branch structure, which allows the model to capture more diverse features and further enhances the adaptability and flexibility of the model.