Improving forecasting accuracy of electric switch demand using artificial neural network and aggregate customer clustering techniques

Demand forecasting has become progressively important for today’s competitive business environment. It affects whole supply chain systems from manufacturing planning to material requirement planning. However, accurate demand forecasting is still difficult and challenging today. The nature of product sometimes creates more difficulties in forecasting for items like electric switches because of their mixed structures. Long lead time, low volume and a large variety of components make the prediction of electric switch demand harder. Traditional forecasting methods are limited to identify the causes of variability in demand and to capture nonlinearity pattern in the demand. However, there are potential alternatives such as artificial neural networks (ANN), which can take into account the external causes that influence the demand variability and analyze nonlinear patterns when it is applied for forecasting. The objective of this research is to improve forecasting accuracy of electric switch demand by using the ANN method. Furthermore, this research investigates cluster-based aggregate forecasting techniques in order to achieve higher accuracy. In the first part of the study, ANN and traditional methods (e.g., moving average, exponential smoothing, Holt’s method and Winter’s method) are applied to forecast demand of two product line of electric switches, D-switch and TMX, and their accessories’ lines, vacuum interrupter and motor operator. The ANN model includes several external factors that influence the switch sales demand. The forecasting results show that ANN has a smaller error rate compared to traditional approaches. In the second part, we focus on interrupter demand and its customer to create meaningful clusters in order to reduce forecasting error. We develop a cluster-based aggregate forecasting (CBAF) model to achieve even higher accuracy by clustering customers into separate groups and applying demand forecasting separately to each cluster, then aggregating them all. The CBAF performs better than forecasting all demand at once with no prior clustering in terms of mean squared error (MSE) and mean absolute deviation (MAD). This study demonstrated that the two proposed methods (ANN and CBAF) have improved the forecasting accuracy of electric switches over traditional methods