Time Series Forecasting Based on Convolution Transformer

For many fields in real life, time series forecasting is essential. Recent studies have shown that Transformer has certain advantages when dealing with such problems, especially long sequence input and problems. In order to improve the efficiency local stability of Transformer, these combine CNN different structures. However, previous network models based on cannot make full use CNN, they not been used a better combination both. response this problem forecasting, we propose algorithm convolution Transformer. (1) ES attention mechanism: Combine external traditional self-attention mechanism through two-branch network, computational cost reduced, higher accuracy obtained. (2) Frequency enhanced block: A Enhanced Block added front ESAttention module, which can capture important structures frequency domain mapping. (3) Causal dilated convolution: The module connected by replacing standard layer causal layer, so it obtains receptive field exponentially growth without increasing calculation consumption. (4) Multi-layer feature fusion: outputs modules are extracted, convolutional layers adjust size map for fusion. more fine-grained information obtained at negligible cost. Experiments world datasets show model structure proposed paper greatly real-time performance current state-of-the-art model, memory costs significantly lower. Compared algorithms, achieved greater improvement both effectiveness accuracy.