中国烟草学报

卷烟市场销量具有强烈的“周期性+节令性+策略性”特征，传统预测模型难以刻画政策指令、供给计划、投放节奏、库存守恒、专卖约束等关键的非市场因素影响。为此，本文将多层感知机（Multi-Layer Perceptron， MLP）与大语言模型（Large Language Model，LLM）相结合，提出了“MLP主导、LLM辅助”的卷烟销量组合预测模型。MLP擅长自然数据建模而LLM具备多源信息理解与推理能力。在此基础上，将销量预测模型应用于卷烟移库策略，提出了基于递阶分布式模型预测控制（Distributed Model Predictive Control， DMPC）的卷烟物流供应链库存动态最优控制方法。采用Z省2022年1月至2023年6月共18个月的真实历史数据进行销量预测和库存控制，结果表明：该预测模型对畅销品规的平均相对预测误差为3.5%，显著优于ARIMA（20.0%）和RBF神经网络（8.3%）；该库存控制方法对某地市前置仓不同时期3项卷烟移库指标（平均库存、运输车次、入库量标准差）均低于传统的存销比控制方法，平均分别下降了24.21%、36.84%和9.93%。

Cigarette market sales exhibit strong cyclical， seasonal， and strategic characteristics， making it difficult for traditional forecasting models to capture the impact of key non-market factors such as policy directives， supply plans， release rhythms， inventory balance， and monopoly regulations. To address this， this paper combines the Multi-Layer Perceptron (MLP) with Large Language Model (LLM) to propose a combined forecasting model for cigarette sales characterized by “MLP dominance and LLM assistance”. In this framework， the MLP excels at modeling natural data patterns， while the LLM provides multi-source information understanding and reasoning capabilities. Based on this forecasting model， the study applies it to cigarette inventory transfer strategies and proposes a dynamic optimal control method for cigarette logistics supply chain inventory using hierarchical Distributed Model Predictive Control (DMPC). Using real historical data from Province Z covering 18 months from January 2022 to June 2023， the sales forecasting and inventory control are evaluated. Results show that the proposed forecasting model achieves a mean relative prediction error of 3.5% for popular stock-keeping units， significantly outperforming ARIMA (20.0%) and the RBF neural network (8.3%). Furthermore， the proposed inventory control method outperforms the traditional inventory-to-sales ratio approach across three cigarette inventory transfer indicators (average inventory， number of transport trips， and standard deviation of warehouse arrivals) in a front warehouse of a prefecture-level city during different periods， achieving average reductions of 24.21%， 36.84%， and 9.93%， respectively.