پیش بینی تقاضای مشتری برای طراحی محصول انتخاب اتوماسیونی و پیکر بندی روشهای پیش بینی مناسب

Demand planning is of significant importance for manufacturing companies since subsequent steps of production planning base on demand forecasts. Major tasks of demand planning are the selection of a prediction method and the configuration of its parameters subject to a given demand evolution. This paper introduces a novel method for the automated selection and configuration of suitable prediction methods for time series of customer demands. The research investigates correlations between dynamic time series characteristics and forecasting accuracies of different prediction methods. The evaluation of the method on a database comprising real industry data confirms excellent prediction results.

Production planning is an important task for manufacturing companies. In particular, demand planning, which is premised on forecasts of future customer demands, is the main basis for all following planning steps [1]. Since no single prediction method outperforms all other methods for all cases of time series [2], two major tasks have to be fulfilled. At first, a suitable prediction method must be selected. Subsequently, the parameters of this method have to be configured to match a current demand evolution. Commonly, manufacturing companies select a few prediction methods, configure their parameters and compare the training results to select the best of the considered methods. In an extreme case, only one prediction method is selected and configured with respect to a current demand evolution. The advantage of this approach is the quick establishment of forecasting results. Nevertheless, the probability of poor results in comparison to other prediction methods is high. The other extreme is to configure various prediction methods to the current demand evolution and to select the method with the lowest training error. This approach can lead to more accurate predictions. However, it requires either a high amount of expert knowledge or automated algorithms to find appropriate parameter configurations of the applied prediction methods. Moreover, it takes much time to configure the different prediction methods and to compare the prediction results. Besides, the increasing number of product variants in recent years further complicates demand planning [3]. The paper at hand presents a novel data-driven approach for the automated selection and configuration of suitable prediction methods. Thereby, sophisticated predictions are calculated quickly. By using machine learning methods, correlations between time series characteristics of demand evolutions and the accuracies of different prediction methods are analysed. Here, measures of recurrence quantification analysis are incorporated in addition to common time series characteristics. Furthermore, the study includes several established prediction methods which model time series evolutions locally or globally as well as linear or nonlinear. After setting up a knowledge base, a suitable prediction method for an unknown time series of customer demands is selected and configured automatically. The remainder of this paper is structured as follows. Section 2 gives a theoretical background. The new automated selection and configuration approach is detailed in Section 3. Section 4 describes an evaluation on real time series data of the M3-competition [4]. Conclusions and an outlook on further research directions are given in Section 5.

In this paper, a new method for the automated selection and configuration of an appropriate prediction method for time series of customer demands was proposed. This method incorporates correlations between 26 time series characteristics and forecasting accuracies of six different prediction methods. An evaluation on a real data set of 200 time series showed excellent prediction results. The proposed method performed better than the six individual methods and a benchmark method using decision trees, less features and less prediction methods. The proposed method also outperformed the more time-consuming method to train all individual methods and to select the one with lowest training error. By using the proposed method, accurate predictions were calculated with low effort of computation time. In further research, other classifiers will be applied and compared to the LDA classifier used in this paper. More individual prediction methods will be included to further improve the predictions.