To improve the accuracy of the Xin’anjiang model（ConvLSTM）integrating attention mechanism was constructed to replace the confluence module in XAJ, thereby establishing a hybrid hydrological model XAJ ACL that combined physical mechanisms and machine learning techniques. The performance of XAJ ACL under the condition of limited sample capacity in small and medium sized watersheds was explored based on the measured data of the Chengcun Watershed. To further evaluated the effect of XAJ ACL, ConvLSTM and the traditional LSTM were respectively adopted to replace the confluence module in XAJ, and two additional hybrid hydrological models, XAJ CL and XAJ LSTM, were constructed for comparative analysis with XAJ ACL. The results show that in runoff simulation of the Chengcun Watershed, the simulation accuracy of XAJ ACL is superior to that of XAJ. During the testing period, the Nash Sutcliffe efficiency coefficient (NSE) of XAJ ACL is 0.85, and the correlation coefficient is 0.93, both of which are higher than those of XAJ. In training of three sets of small capacity samples, the average NSE values of XAJ ACL during the testing period were 0.847, 0.832, and 0.808, respectively, all higher than those of XAJ CL and XAJ LSTM, and the simulation results showed better stability. Compared with XAJ, XAJ ACL significantly improves the ability for simulation of nonlinear laws of confluence processes in small and medium sized watersheds under limited data conditions.