Abstract: In the light of multiple problems in real-time battery capacity prediction of energy storage system,such as the different utilization degree of batteries and the various sampling intervals of discharge rate,a new short and long time memory network based on ampere-hour residual connection is proposed. The characteristic transformation of multi-channel sensor signal makes it more suitable for real-time residual capacity prediction of diversified battery discharge. Compared with common baseline methods,the error of proposed method is reduced by 11.8%. The proposed method has smaller variance and higher robustness under multiple rate test.