Abstract:

[Objective] For the tobacco industry, it is crucial to improve the commodity resource allocation strategy of retailers, optimize the supply policy of commodities, and solve the problems of unreasonable allocation of goods, behindhand commodities supply policy and low efficiency of business operations, which exist in the current commodities allocation based on the historical sales level and experiences but lack of systematic data theory support [Methods] A multi-dimensional data-driven neural network model is proposed. the data used by the model includes merchant internal order data and external shopping area, etc. By abnormal sample removal, feature construction, and partition of training and validation data sets, the attention layer, the dense layer, and the LightGBM model layer are passed successively on the training set. [Results] A prediction accuracy of 96.57% is finally achieved on the test set. [Conclusion] Based on the multi-dimensional data of the cigarette system, this technology can establish a highly adaptable and flexible demand forecasting system, break the current operation mode of determining the demand according to the stalls, realize high prediction accuracy, and provide technical support for the intelligent placement of cigarettes for enterprises.

Key words: neural networks, demand forecasting, feature construction, Attention layer, LightGBM model