Abstract: In recent years,time series forecasting has found widespread applications in various domains such as finance,meteoro-logy,and military.Deep learning has begun to demonstrate significant potential and application prospects in time series forecasting tasks.However,recurrent neural networks often encounter issues like information loss and exploding gradients when dealing with time series predictions over extended periods.In contrast,Transformer models and their variants,when utilizing attention mechanisms,typically overlook the temporal relationships between variables in time series data.To address these challenges,this paper proposes a hybrid neural network time series forecasting model based on seasonal decomposition.This model employs a seasonal decomposition module to capture the variations in different periodic frequency components within the time series.Simultaneously,by integrating multi-head self-attention mechanisms and composite dilated convolution layers,the model leverages the interaction between global and local information to obtain multi-scale temporal positional information among the data.Ultimately,experiments are conducted on publicly available datasets from 4 different domains,and the results indicate that the predictive perfor-mance of the proposed model surpasses that of current popular mainstream methods.

Key words: Time series forecasting, Seasonal decomposition, Self-attention mechanism, Dilated convolution, Hybrid model