A Hybrid Data-Driven Approach for Optimizing Last-Mile Logistics in an Electrical Products Company Using K-Means Clustering and the Clarke-Wright Savings Algorithm

Abstract

Managing last-mile delivery presents significant challenges when customer locations are widely dispersed, route demand fluctuates across dispatch days, and planning relies on historical route groupings. This study introduces a hybrid, data-driven decision-support framework designed to enhance last-mile delivery planning for an electrical products company. The baseline system was evaluated using 2025 delivery records and operational indicators, including distance per trip, cost per drop, backload rate, loaded volume, drops per trip, trip duration, and truck utilization. Pareto analysis identified recurring backload categories that impeded delivery completion. The proposed framework integrates K-Means clustering with the Clarke-Wright Savings Algorithm. K-Means assigned 1,573 customer delivery points to five geographically coherent service clusters based on coordinate proximity. Within each cluster, the Clarke-Wright algorithm established preliminary route groups by prioritizing customer pairings with greater distance savings. Route outputs were validated against operational constraints, such as a 4-CBM capacity threshold, route continuity, loop prevention, a practical stop-count range of 17 to 22 drops, and flexible merging rules for low-density groups. The framework produced 79 preliminary route groups, all of which satisfied the 4-CBM capacity constraint, while 69 groups (87.34%) met the practical stop-count range. Simulation-based comparisons demonstrated distance reductions ranging from 37.77% to 77.59% across comparable clusters. A controlled-cost simulation for 4W deliveries indicated that the optimized route structure could reduce estimated distance-related fuel costs and decrease the cost per drop from PHP 132.45 to PHP 114.03 under constant-cost assumptions. These findings support the framework as a practical reference for route planning, pending pilot validation using road-network distance, traffic conditions, unloading time, customer receiving conditions, and actual dispatch records.