[2H4-E-2] Machine learning: fusion of models

[2H4-E-2-02] Improvement of Product Shipment Forecast based on LSTM Incorporating On-Site Knowledge as Residual Connection

〇Takeshi Morinibu¹, Tomohiro Noda¹, Shota Tanaka¹ (1. Daikin Industries, Ltd. Technology and Innovation Center)

Keywords:Product Shipment Forecast, Supply Chain Management, LSTM, ResNet, ARIMA

It is important to predict shipments of air conditioners for the purpose of making a production plan. Although ARIMA was used for that prediction for a long time, it turned out that some products we manage had less accurate prediction score. In order to get more precise prediction, we applied LSTM to forecast shipments. Despite the complexity of LSTM, we could not get what we expected. Therefore, we further improved the accuracy by adding on-site knowledge to network structure of LSTM as residual mechanism.