即时配送履约效率的提升依赖对时长形成机制的精准刻画，而传统人工规则与单一统计模型难以充分覆盖空间轨迹异质性、骑手个体差异与长尾分布的复杂影响。既有研究多侧重模型精度提升，对时空特征的增量效度、多模型统一比较标准与误差场景分布的讨论相对不足。本研究基于31415条即时配送订单轨迹数据，构建包含三类Haversine球面距离的时空行为特征体系，在统一训练测试划分与5折交叉验证框架下对比 10 类回归模型的预测表现，并借助 SHAP 方法解析核心影响因素的作用机制。研究发现：时空行为特征可显著提升基础模型的预测精度，其增量效度稳定存在于不同模型范式中；树集成模型整体优于线性模型与神经网络基线，随机森林取得最优OOF表现，RMSE为88.07、MAE为57.10、R2为0.6128；模型误差随配送时长增加而扩大，在高定位偏差、跨区域履约与高峰时段等场景下更为集中；收货点到送达点距离、骑手个体差异、时间编码与全链路空间距离是影响时长预测的核心因素。研究表明，将配送业务过程转化为可解释的时空特征，结合集成学习与 SHAP 解释框架，能够同时兼顾预测精度与业务可读性，为即时配送精细化运营提供方法支撑。

The improvement of instant delivery fulfillment efficiency relies on the precise characterization of the duration formation mechanism. However, traditional manual rules and single statistical models struggle to fully cover the complex impacts of spatial trajectory heterogeneity, individual differences among riders, and long-tail distribution. Existing research often focuses on improving model accuracy, yet discussions on the incremental validity of spatiotemporal features, unified comparison standards across multiple models, and error scenario distribution are relatively lacking. Based on trajectory data from 31,415 instant delivery orders, this study constructs a spatiotemporal behavior feature system encompassing three types of Haversine spherical distances. It compares the predictive performance of 10 regression models under a unified training-testing split and a 5-fold cross-validation framework. Furthermore, it utilizes the SHAP method to analyze the mechanism of core influencing factors. The study finds that spatiotemporal behavior features can significantly enhance the predictive accuracy of the basic model, and their incremental validity consistently exists across different model paradigms. The tree ensemble model outperforms the linear model and neural network baseline overall, with random forest achieving the best OOF performance, with an RMSE of 88.07, MAE of 57.10, and R² of 0.6128. Model errors increase with the duration of delivery, and are more concentrated in scenarios such as high positioning deviation, cross-regional fulfillment, and peak hours. The distance from the pickup point to the delivery point, individual differences among riders, time encoding, and the spatial distance of the entire route are core factors affecting duration prediction. The research indicates that transforming the delivery business process into interpretable spatiotemporal features, combined with ensemble learning and the SHAP interpretation framework, can simultaneously achieve both predictive accuracy and business readability, providing methodological support for refined operations in instant delivery.