Precisely estimating the battery State of Charge (SOC) can maximize the utilization of battery energy and reduce energy consumption. The tradition neural network-based battery SOC estimation methods have a common defect, that is, the input of the model is a continuous constant current or a periodic current that does not change drastically under normal power. These traditional methods are not accurate when the battery is in a low power condition. Aiming at the above problems, this paper proposed a battery SOC estimation method based on a gated recurrent unit recurrent neural network with self-attention mechanism (GRU-Attention) under low power. Firstly, to validate the new method, experiments were conducted using datasets from real battery tests. Then, preprocess the data by serializing the discharge data of the battery. Finally, these data are input into the GRU-Attention model for training. The experimental results indicate that under low power, the maximum error for DST dataset is 4.75%, while for UDDS dataset, it is only 3%, demonstrating its capability to accurately estimate battery SOC. In summary, the new approach provides a novel and effective pathway for battery management under low power conditions, holding significant theoretical and practical implications for the advancement of battery technology.