关键词: 汽车配件, 多价值链, 轻梯度提升机, 门控循环神经网络, 优势矩阵, 组合预测模型, 半组合预测模型, 子模型优选, 需求预测

Abstract: Based on the sales status of auto parts in the third-party parts multi-industry chain business collaborative cloud service platform,parts agents have not fully considered the parts demand of cross-chain sales,cross-chain allocation,and multi-chain sales.To improve the accuracy of parts demand forecasting,a combination prediction model combining Advantage Matrix (AM) with Light Gradient Boosting Machine (LightGBM) and Gated Recurrent Unit (GRU) named LightGBM_GRU_AM was proposed.The model obtained the optimal weight coefficient of a single model by introducing the advantage matrix and forecasted the demand through the weighted combination model.Considering that the prediction effect of the sub-model at a certain moment was better than the combined model,a semi-combined prediction model based on LightGBM,GRU,and LightGBM_GRU_AM was designed to further improve the accuracy of the prediction,which adopted the sub-model optimization strategy.In the training process,the minimum absolute error was used to establish sub-model classification labels,the feature extraction and classification regression trees were used to establish sub-model selection rules,the different sub-models were used to make prediction according to data characteristics,and the predicted values ??at different times were integrated to form the final demand forecast.The multi-chain parts sales and parts-related after-sales service data in the integrated the third-party cloud platform were analyzed by calculation examples.Compared with the other seven forecasting models,two proposed forecasting models could effectively reduce the forecast error,and the semi-combined forecasting model had more advantages,which could provide purchasing decision support for parts agents.

Key words: auto parts, multi-value chain, light gradient boosting machine, gated recurrent unit, advantage matrix, combination forecasting model, semi-combination forecasting model, sub-model selection, demand forecasting