Dynamic routing optimization of electric vehicles for retailers based on consumer behavior prediction

Abstract

To address the dual challenges of declining retailer profit margins and rising logistics costs, we propose to use the surplus capacity of electric vehicles (EVs) to conduct opportunistic sales and construct a new decision-making framework that integrates consumer behavior prediction and a dynamic customer insertion strategy to improve corporate profit margins by actively exploring potential demand. First, a stacked ensemble learning model is built based on user historical behavior data to predict the user's purchase probability, accurately identify high-purchase probability customers and push demand product orders to them. Then, a two-stage dynamic path optimization model is designed. Static customer order information and prediction probability are used as input parameters of the model for path optimization in the initial stage. By setting a high-probability purchase customer threshold, high-purchase rate customers are screened out, and resources are pre-allocated. In order to quantify the prediction uncertainty, the Chernoff approximation is introduced to transform random constraints into deterministic constraints, and the safety margin is calculated to ensure that the pre-allocated resources can meet the dynamic demand. In the second stage, combined with the improved Bernoulli distribution, the actual dynamic customer demand is calculated based on the prediction probability, prediction accuracy and deviation factor. The path re-optimization is achieved through the profit-maximizing insertion strategy, and the battery replacement solution is introduced to reduce the risk of battery life interruption. In terms of solution algorithm, we adopt the mileage saving-tabu search hybrid algorithm, integrate multiple local search operations, and enhance the global optimization capability. Experimental results show that: (1) The pre-allocation strategy based on the prediction results can make the service rate of high-probability purchase customers reach up to 100%, and the enterprise profit increases by 30.598%. (2) The average solution time of the solution strategy we proposed is 5.7382 s, the average cost increase rate is 21.1975%, and the average increase in cost per dynamic customer is 26.1495 CNY. These three key indicators are significantly better than the two strategies of rolling time domain and pre-processing all dynamic requirements in advance.