Previsão de demanda de autopeças com redes neurais


IBICT - Instituto Brasileiro de Informação em Ciência e Tecnologia




This paper presents a methodology for forecasting demand parts based on Artificial Neural Networks (ANN). To validate it, we performed a comparative study on a reference work in the literature, which is based on exponential smoothing and moving average methods. The products are grouped into 10 categories according to proximity, resulting on 72 monthly observations. The forecasts generated were qualitatively classified according the criterion of lower mean absolute percentage error (MAPE), and the ANN model showed potential capability on 70% of instances considered, with respect to the best models treated by the author, and effective capability on 60%of the total. Qualitatively, the only two instances that showed the best ratings on a scale of four classes have been implemented with the neural networks based model. This methodology was then applied to a case study about a company of national operation in the auto parts market, with stock of more than 11,500 different products. In this application, a new approach to temporal grouping was proposed, different of the weekly and monthly format, plus the unbundling of products in categories in order to avoid the influences arising out of this type of aggregation and solve the problems caused by intermittent daily demand. Data of demand were collected from January 2007 to July 2009, totaling 122 values in temporal clustering defined as "fourth of month." The sample selection for this case study, conducted by Pareto classification, showed that just over 12% of total products were responsible for more than 80% of monthly volume sold by the company. Among other steps, this paper proposed the treatment of outliers of the time series through the transformation of the instances, the analysis of autocorrelation of the original and transformed series, and analysis qualitative results of the forecasts. It was observed that the methodology based on ANNs outperformed qualitatively the most of the results of reference methods literature, both the comparative approach as in the case study


previsão de demanda redes neurais ciencia da computacao

Documentos Relacionados