In response to the challenges of high computational cost associated with traditional physics-based models and the poor generalization of purely data-driven methods in predicting sedimentation in port basin channels,a hybrid sediment transport prediction framework that integrates physical principles with sparse data-driven learning is proposed.The physics-guided long short-term memory (PhyLSTM) model built in this paper is incorporated key physical parameters from sediment transport-diffusion equation and bed evolution model (such as the diffusion coefficient and source-sink term,etc.)as learnable variables.The bed elevation trend is also embedded as a physical prior to guide the learning of nonlinear sediment dynamics.Using 1,440 sets of multi-day monitoring data collected over three years from a typical port basin (0.002 5°×0.002 5°grid),the prediction performance of PhyLSTM is compared with a traditional LSTM model.The results show that the PhyLSTM model achieves a mean absolute error of 0.091 5 m,a root mean square error of 0.114 5 m,and a mean absolute percentage error of 0.75%,representing reductions of 37%,30%,and 36% with the traditional LSTM model,respectively.Additionally,91.74% of predictions fall within a ±0.2 m error range,10.18% higher than that of the LSTM.The proposed method offers a novel,interpretable,and accurate approach for predicting sedimentation in port basin channels,which holds substantial engineering value for optimizing port dredging strategies and minimizing maintenance expenses.