MDWConv:CNN based on multi-scale atrous pyramid and depthwise separable convolution for long time series forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Long time series forecasting is crucial for applications like power dispatching and weather prediction.
• Transformer-based models dominate but struggle with multi-scale and inter-variable information.
• Existing methods often lack efficient handling of complex temporal dependencies.

Purpose of the Study:

• To propose a novel convolutional neural network (MDWConv) for enhanced long time series forecasting.
• To address limitations in capturing multi-scale features and inter-variable interactions.
• To improve predictive accuracy and computational efficiency.

Main Methods:

• Developed MDWConv, a convolutional neural network utilizing a multi-scale dilated pyramid for feature extraction.
• Employed depthwise separable convolution with a grouping strategy for long-term dependencies and inter-variable interaction.
• Introduced a novel segmented polynomial activation function (TCP) to approximate GELU, reducing computational load.

Main Results:

• MDWConv demonstrates superior performance compared to existing methods on various real-world datasets.
• The model effectively integrates multi-scale information and captures inter-variable dependencies.
• The proposed TCP activation function accelerates loss reduction and decreases computational complexity.

Conclusions:

• MDWConv offers a competitive and efficient alternative for long time series forecasting using solely convolutional neural networks.
• The architecture successfully handles multi-scale and inter-variable information, crucial for complex forecasting tasks.
• MDWConv presents a promising direction for advancing time series analysis.