At present, the data-driven based SOC estimation methods require the collection of a substantial quantity of datasets for the purpose of complex training. This paper proposes a transfer learning SOC estimation method based on Long Short Term Memory (LSTM) with self-attention mechanism. The objective is to provide accurate SOC estimation for small target sample real vehicle datasets. Firstly, learn rich voltage and current features by processing long time series with LSTM. Subsequently, the lengthy temporal span input sequences are compensated with the self-attention mechanism, which results in a global dependency model. Combine the model with transfer learning to solve problems in the target domain using knowledge learnt from the source domain. The experimental results demonstrate that the proposed method exhibits superior SOC estimation accuracy and shorter training time, and verifies the effectiveness for small sample real vehicle datasets.