Surface turbulent fluxes are a key energy exchange in the atmospheric boundary layer (ABL). Accurately predicting its variations is essential for agricultural ecology, and climate studies. Current prediction methods include the Monin-Obukhov Similarity Theory (MOST) and Machine Learning (ML). However, MOST need experimental parameters and empirical formulas, and ML relies much on manual feature extraction. Given the potential of Deep Learning (DL) in time series prediction, this study employs the Inverted Transformer (iTransformer) to forecast friction velocity, kinematic sensible heat flux, and kinematic latent heat flux across different seasons. iTransformer encodes the data using transposed encoding and utilizes a multi-variate self-attention to capture the correlations between variables. The feed-forward neural networks leverage these correlations to predict the surface turbulent fluxes. Compared with other methods including Transformer and ML methods, the iTransformer model can not only improve the prediction correlation but also reduce errors of surface turbulent fluxes. Meanwhile, the model can effectively capture the fluctuation trends of various fluxes within a month or even a day. In conclusion, the iTransformer can significantly improve the predictive performance of surface turbulent fluxes.