In order to address the challenges of long-and short-term dependencies in time series forecasting，while capturing both long-term trends and short-term dynamics，and to model the complex interdependencies among variables in multivariate time series，a forecasting method based on the temporal convolutional network （TCN） was proposed. Initially，the TCN was employed to effectively capture the features of sequence variables across time scales，and subsequently，the squeeze-and-excitation block （SE_Block） was applied to the outputs of the TCN. This module was used to enhance the representation of multivariate variables，resolve short-term dependency issues，and improve the model’s ability to capture critical short-term information. Secondly，the Informer model was introduced to enhance the long-sequence processing capability，which not only effectively addressed the computational efficiency issue in long-sequence forecasting but also strengthened the model’s ability to capture global temporal dependencies. Finally，the proposed method was experimentally validated and compared with existing time series models on three datasets： equipment state monitoring（ETTm1），traffic flow （Traffic），and electricity load （Electricity）. The results show that the proposed method exhibits lower error rates in both long-and short-term time series forecasting. It is concluded that the proposed method can effectively improve the forecasting performance for multivariate time series in both long-and short-term domains.