In order to improve the robustness and link prediction accuracy of wireless communication in smart grid, a low confidence interval prediction method for link quality of smart grid based on long-term and short-term memory network is proposed. Firstly, the communication requirements of smart grid and the characteristics of wireless link are analyzed. According to the analysis results, the signal-to-noise ratio time series is decomposed into deterministic part and stochastic part by wavelet denoising algorithm. On the one hand, the deterministic part is input into the two-layer long-term and short-term memory neural network for prediction. On the other hand, the variance time series of the random part is calculated and then used as the input of another two-layer long-term and short-term memory neural network for prediction. Then the variance of the deterministic part and the stochastic part are predicted respectively, and the confidence interval boundary is calculated. Experimental results show that the proposed method has better robustness and prediction accuracy.