As one of the most important functions of the battery management system, state of charge (SOC) estimation has been widely concerned since it is crucial to the operating safety and reliability of lithium-ion batteries. In this paper, an attentional sequence-to-sequence network is proposed to achieve accurate SOC estimation based on voltage, current, and temperature measurements directly. An attention mechanism is introduced to the network to capture long-term dependency and to reduce the loss of context information in measurement sequences, thus making full use of context information from temporal measurement sequences. The proposed method is trained using a part of public battery datasets, which involve different driving cycles at −20 °C, −10 °C, 0 °C, 10 °C, and 25 °C. The performance of the network is validated by the rest of the public battery datasets, including both different fixed temperatures and continuously varying temperatures, and datasets collected from different battery types. Experimental results reveal that the proposed method achieves robust and accurate SOC estimation under different driving cycles which are not contained in the training datasets, and the root mean square error is within 3.1 %. Even for continuously varying temperatures and different battery types, the root mean square error of the proposed network for SOC estimation is still lower than 3.8 %.