AMDCnet: attention-gate-based multi-scale decomposition and collaboration network for long-term time series forecasting

Time series analysis plays a critical role in various applications, including sensor data monitoring, weather forecasting, economic predictions, and network traffic management. While traditional methods primarily focus on modeling time series data at a single temporal scale and achieve notable results, they often overlook dependencies across multiple scales. Furthermore, the intricate structure of multi-scale time series complicates the effective extraction of features at different temporal resolutions.

To address these limitations, we propose AMDCnet, a multi-scale-based time series decomposition and collaboration network designed to enhance the model's capacity for decomposing and integrating data across varying time scales. Specifically, AMDCnet transforms the original time series into multiple temporal resolutions and conducts multi-scale feature decomposition while preserving the overall temporal dynamics. By extracting features from downsampled sequences and integrating multi-resolution features through attention-gated co-training mechanisms, AMDCnet enables efficient modeling of complex time series data.

AMDCnet achieving 44 best results and 10 second-best results out of 64 cases. Experimental results on 8 benchmark datasets demonstrate that AMDCnet achieves state-of-the-art performance in time series forecasting.

Our research provides a robust baseline for the application of artificial intelligence in multivariate time series forecasting. This work leverages multi-scale time series decomposition and gated units for feature fusion, effectively capturing dependencies across different temporal scales. Future studies may further optimize the scale decomposition and fusion modules. Such efforts could enhance the representation of multi-scale information and help address key challenges in multivariate time series prediction.