基于集成学习的物资采购价格辅助决策方法  

Auxiliary Decision Making Method of Material Purchase Price Based on Ensemble Learning

作　　者：程晓晓 蒲兵舰 张国平 丁萌萌 CHENG Xiaoxiao;PU Bingjian;ZHANG Guoping;DING Mengmeng(Department of Procurement,State Grid Henan Electric Power Company,Zhengzhou 450000,China;Materials Company,State Grid Henan Electric Power Company,Zhengzhou 450000,China)

机构地区：[1]国网河南省电力公司物资部,郑州450000 [2]国网河南省电力公司物资公司,郑州450000

出　　处：《吉林大学学报（信息科学版）》2022年第5期875-883,共9页Journal of Jilin University（Information Science Edition）

基　　金：国网公司总部基金资助项目(5400-202124146A-0-0-00);国网河南省电力公司基金资助项目(5217N0210001)。

摘　　要：电力所需物资种类繁多且物资价格的波动受到多种因素的影响,为预测当下价格走势,建立电缆价格预测模型,为电网公司提供招标底价的依据与合理采购意见,对收集得到的物资价格,利用动态时间规整方法确定物资的不含税单价滞后于原材料价格的时间,从而确定不含税单价、原材料价格、经济指标间的对应关系。影响物资价格变化的因素很多,利用皮尔逊系数和随机森林两种方法筛选得到关键特征。根据选定的关键特征和数据分别建立AdaBoost、 XGBoost(Extreme Gradient Boosting)、随机森林3种模型对物资价格进行预测。利用预测评价指标平均绝对百分比误差(MAPE:Mean Absolute Percentage Error)评估预测效果,结果表明利用随机森林筛选关键特征配合XGBoost模型进行预测的准确率最高。Power companies need many kinds of materials, and the fluctuation of material prices is affected by many factors. To predict the current price trend, a cable price prediction model is established which can provide the basis of bidding base price and reasonable procurement opinions for power companies is eslablished. The dynamic time warping method is used to determine the price delay time for the collected material prices, that is, the time when the tax free unit price of materials lags behind the price of raw materials. And the corresponding relationship among unit price excluding tax, raw material price and economic indicators is determined. There are many factors affecting the change of material price. The key characteristics are screened by Pearson coefficient and random forest. According to the selected key characteristics and data, AdaBoost, xgboost and random forest models are established to predict the material price. The average absolute percentage error of the prediction evaluation index is used to evaluate the prediction effect. It is found that the key features screened by random forest and then combined with XGBoost model have the highest accuracy of prediction.

关 键 词：价格预测 集成学习 XGBoost模型 皮尔逊系数 随机森林 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] F726[自动化与计算机技术—控制科学与工程]