DS-ALNS for Practical Sorting Optimization in Parcel Distribution Centers

Parcel distribution centers (PDCs) of major carriers aim to efficiently sort the inbound parcels into downstream destinations via a parcel-sorting system (PSS). Inbound parcels are released into the PSS and diverted to the assigned grids, followed by packing into bins and loading onto outbound trucks. Maximizing the sorting throughput requires an optimized assignment of parcel destinations to grids, which is considerably complex when incorporating the dynamic models of parcel movement, packing behaviors, conveyor system congestion, and the practical operation constraints posed by real-world sorting. Given the limited research on the sorting optimization of the cross-belt sorter system (CBSS), we first provide a systematic introduction to its structure, and conduct detailed throughput analysis. Sub-sequently, we establish a specific optimization formulation with practical operational constraints. Then, we formulate the sorting optimization problem into a uniformed grids assignment problem and decompose it into the grids quantities assignment and constraint-based positioning assignment problems, circumventing the computational challenge posed by the enormous solution space and constraint dimensions while ensuring operational feasibility. Targeting each of the sub-problems, we propose a dual-stage adaptive large neighborhood search (DS-ALNS) scheme, which implements the ALNS algorithm to tackle the grid quantities and constraint-based positioning assignments, respectively. Moreover, we implement a rank correlation transformation design to tackle the computational complexity challenge by transforming the decision variables from determining the assigned quantities of the grids into the destinations, significantly reducing the solution space. We evaluate the effectiveness and operational feasibility through the Digital Twin (DT) system in various PDCs with real-world sorting data. The DT experiments results demonstrate that the proposed DS-ALNS can achieve a decent increase in both stationary and peak sorting throughput up to 3.08% and 7.42% respectively, underlining the effectiveness and practical applicability.