Long Time Series Forecasting Based on Dilated Causal Convolution and Focused Linear Attention with Informer

There are many time series data in our daily life and accurate prediction of these data can bring great convenience to our life. In this study, a hybrid prediction model based on dilated causal convolution, focused linear attention and Informer is proposed to improve the accuracy of predicting long time series. There are two main innovations in this paper:(1) Dilated causal convolution is added to the Informer model to extract the time series features and increase the sensory field of the model. (2) Focused linear attention is proposed to reduce the time complexity of predicting long sequences. The proposed model exhibits a 33.3% improvement in MSE metrics on the WTH dataset in comparison to Informer, an 89.4% improvement relative to Autofomer, a 18.9% improvement relative to Transformer, a 17.4% improvement relative to Pyraformer, and a 30.7% improvement relative to TimesNet. Consequently, the model is positioned as a frontrunner in the domain of time series prediction. In addition, the hybrid model demonstrates favorable outcomes on datasets pertaining to traffic, WTH, and other domains.