Abstract: Aimed at the problem of insufficient accuracy of short- distance cargo volume prediction and weak interpretability of the model, on the basis of considering the high- dimensional feature interaction and hyperparameter tuning cost, a LightGBM short- distance cargo volume prediction method (SHAP- TBO- LightGBM) based on improved Bayesian optimization is proposed. Bayesian optimization was triggered when the drift rate alarm index was triggered, otherwise the historical optimal parameters were reused. The optimal LightGBM was used to predict the total volume of goods as a result on December 16 on 140 station data from both shipping nodes. The experimental results show that the number of retraining times is reduced by 75%, the average R^2 of 140 stations of the two shipping nodes reaches 0.96, and the MASE decreases to 0.18 by using SHAP- TBO- LightGBM, which verifies its practical value in the precise transportation of short- distance freight and provides new ideas and directions for logistics volume forecasting.