基于组合算法的电子产品回收预测系统研究

Electronic products returns forecasting system based on a hybrid algorithm

对第三方逆向物流服务商而言,电子产品回收数量具有少样本、不确定性及模糊性的特点,电子产品回收量预测的精度直接影响到企业的运营成本以及服务水平。在单个预测模型中,GM(1,1)模型具有适应少样本预测的特点,对近期数据具有较好的逼近效果,但是对序列的趋势性比较敏感;FTS模型能够处理不确定性数据中因模糊性而产生的噪声,但是对序列趋势的把握具有滞后性。本文设计了GM(1,1)模型与FTS模型相结合的组合预测模型(FTS_GM(1,1)模型),通过利用两个模型的优势以提高电子产品回收预测的准确性和可靠性。本文根据企业的真实回收数据进行预测,实验结果表明组合预测法比单个预测法具有更好的预测效果。在此基础上,本文提出了以FTS_GM(1,1)组合模型为主,其他预测模型为辅的回收预测系统原型,为企业在实践中选取合适的预测模型提供建议。

The electronic products have special characteristics such as various categories, short lifecycle, and stochastic sales. It is becoming difficult for the third reverse service provider to forecast the quantity and quality of returns accurately because of the uncertainties of consuming circumstance, using habit and location of recovery, the reverse logistics of electronic products encounter the uncertainties of collecting quantity, collecting time and collecting quality. For the third reverse logistics service provider, improving the accuracy of returns forecasting is crucial to improve its operational efficiency and service quality. Firstly, this paper introduces the single forecasting models, including GM(1,1), RGM(1,1) and FTS, and analyzes their characteristics and applicable conditions theoretically. The GM(1,1) model has good approximation effect on the latest data, but it is sensitive to the trend of the sequence. The fluctuation of the sequence will greatly affect the forecasting performance. The FTS model can deal with the noise caused by fuzziness in uncertain data and mine more information in the original sequence. It has better adaptability to the fluctuating sequence. However, there will be a certain lag in grasping the trend of the sequence. This paper considers the uncertainty and fuzziness of the quantity of the returns, and proposes a two-period forecasting model FTS_GM(1,1) based on GM(1,1) and FTS to forecast better by utilizing the advantages of each model. In the first period, when the development coefficient of GM(1,1) |a| ≤1, the forecasting value in GM(1,1) is adapted as the forecasting value in FTS_GM(1,1) because of the good approximation effect on the latest data of GM(1,1). After period one(in the second period), the FTS model is introduced because the error of GM(1,1) will increase greatly with the exponential growth of grey interval with the gradual departure from the actual sequence. The forecasting value in FTS_GM(1,1) is the weighted value of the ones in the GM(1,1) and FTS with variable weights. Due to the worse performance of GM(1,1) in the long term, the variable weight of GM(1,1) is influenced both by the initial weight and the weight attenuation coefficient. In the numerical experiment, this paper collects the historical data of the quantities of the returns in a reverse logistics firm. At first, four forecasting models(GM(1,1)、RGM(1,1)、FTS、FTS_GM(1,1)) are used to make a forecast for one kind of electronic product in the firm respectively, and the results show that FTS_GM(1,1) performs the best. To examine the applicability of FTS_GM(1,1) for other products, five models(moving average model, GM(1,1), RGM(1,1), FTS, FTS_GM(1,1)) are used to make a forecast for all products in the firm respectively. The results also show that FTS_GM(1,1) performs the best in each indicator(MAD、MSE、MAPE) and there is still a lot of room for the performance improvement of moving average method applied in the firm due to the limitation of its management decision level. Also, it is not difficult to find from the comparison results that no method will always perform the best for all types of returns because the sequence of returns quantity is various in an uncertain system due to the complex and changeable situation in practice. Each forecasting model has its limitation, and a single model cannot be applied to all products. Finally, this paper proposes the returns forecasting system prototype based on FTS_GM(1,1) supplemented with other forecasting models, which can examine the selected models in practice and match the best forecasting model, to improve forecasting performance for different products. The reverse logistics service provider can improve forecasting performance by selecting the appropriate model in return forecasting system and reduce the risk of over or insufficient purchase of maintenance spare parts. By this way, the service provider will reduce the spare part inventory and accelerate responding speed to improve its operational efficiency and service quality.