Abstract: Currently,in the research of time series forecasting tasks,Transformer-based models primarily focus on extracting global and local features from time series data and improving attention mechanisms to reduce model complexity.However,exis-ting methods often overlook the different granularity features exhibited by time series at multiple scales.To address this issue,this paper proposes a time series forecasting model that integrates multi-scale features and the attention mechanism,called MTSformer.Firstly,by down-sampling the original sequence,multiple scale subsequences are obtained,enabling the model to integrate multi-scale feature information and enhance generalization ability.Then,a multi-prediction head structure is used to replace the traditional decoder,which improves prediction speed while reducing model complexity.Finally,experiments are conducted on five benchmark datasets,and the results show that compared with existing methods,the MTSformer model achieves ave-rage reductions of 24.51% in MSE and 17.84% in MAE for time series forecasting.

Key words: Time series forecasting, Multi-scale features, Transformer, Multi-head prediction, Downsampling