Multi-Source Domain Adaptation Enhanced Warehouse Dwell Time Prediction

Warehouse dwell time (WDT) of a truck is a critical metric for evaluating plant-logistics efficiency, including the time of the truck's queuing outside and loading inside the warehouse. But WDT prediction is challenging as it is affected by diverse factors like loading distinct types and weights of the cargoes, and varying amounts of loading tasks in different time slots. Besides, each trucks' WDT is transitively influenced by its preceding trucks' loading time in the queue. In this paper, we propose a multi-block dwell time prediction framework consisting of LSTM model and self-attention mechanism, called SDP. In view of that low performance of SDP brought by sparse loading data of some warehouses, we further design a multi-source adaptation based block-to-block transfer learning module. We present a warehouse similarity measurement based on loading tasks allocated and loading ability of the warehouses, according to which we enhance overall prediction performance by learning from high-performance WDT prediction models of similar warehouses. Experimental results on a large-scale logistics data set demonstrate that our proposal can reduce Mean Absolute Percentage Error (MAPE) by an average of 10.0%, Mean Absolute Error(MAE) by an average of 16.5%, and Root Mean Square Error(RMSE) by an average of 17.0% as compared to the baselines.