Last Mile Delivery (LMD) relates to all product delivery activities to the end customer in the supply chain. Timely delivery of products to customers is one of the primary objectives of businesses rendering LMD services. LPG cylinder distribution in India is one such business that provides LMD to its subscribed customers through delivery agents employed by distributors. As per the existing system, the cylinders are brought to the office premises from the warehouse using Delivery Trucks (DT) and are delivered to customers using small delivery vehicles (SDV). The orders delivered each day by the SDV exceed their stock holding capacity. So, SDV must reach the office location from the last delivered customer’s location to restock. This current practice of restocking increases the overall distance travelled by SDV and increases. Further, this increases the delivery lead time, which directly affects customer satisfaction. To overcome these logistics issues in the existing system, this study proposes a dynamic operational location (DOL) for stationing each DT around the area, where the delivery is going to happen each day, considering the customers ordering pattern, instead of all the DT stationed at the office premise. To prescribe DOL for stationing each of the DT this study proposes a hybrid approach and does efficient LMD of LPG cylinders.

The proposed hybrid approach involves a multi-step process. In the first step, a forecasting model is developed to analyze the customer’s order interval pattern and predict their next ordering date range (i.e., the next order interval). Customers whose next ordering date range is within the next ten days from the decision-making epoch/date are segregated and clustered in the second step. The total number of clusters equals the total number of DT operated. In the final step, the centroid location of the generated cluster is prescribed as the DOL to station a DT. Based on the proposed DOL for each DT, the SDV will stock the cylinders from the DT and proceed for LMD.

The proposed hybrid approach is demonstrated using real-life data from a large-scale LPG cylinder distributor in a metro city in Tamilnadu. Accordingly, the dataset used for this study has the customer’s unique consumer numbers, geocoded addresses, and order placement dates for the past five years. Using data pre-processing and feature engineering techniques, the ordering intervals for each customer are obtained, and a forecasting model is developed to predict the next order interval for each customer. Customers’ geolocations with predicted ordering date range within the next ten days are clustered using the K-Means clustering algorithm, and their centroid location is prescribed as DOL for each DT. Finally, both existing practice and the proposed hybrid approach for LMD are compared in terms of the distance travelled by an SDV in the current system with the proposed system.