Cluster-based industrial KPIs forecasting considering the periodicity and holiday effect using LSTM network and MSVR

Abstract

Key performance indicators (KPIs) are key indicators to measure the status and performance of various facilities, time, and resources in the process of production and life. Forecasting industrial KPIs helps in analyzing the future state to make better production and scheduling decisions. However, due to the complexity of industrial production activities, forecasting industrial KPIs becomes an important and challenging task. In this paper, considering the periodicity and holiday effect of data, a cluster-based industrial KPIs forecasting method using long short-term memory (LSTM) network and multi-output support vector regression (MSVR) is proposed. First, based on the knowledge of production and scheduling, the original daily data is divided into workdays, off-days, and holidays, and K-shape clustering divides holidays into multiple sub-series according to the characteristics of distinct holidays. Then, for the new prediction task, the most suitable historical sub-series is selected through a matching strategy combined with date information. Finally, a compatible model is adaptively selected for forecasting according to the type of sub-series. Specifically, LSTM network is employed to forecast the KPIs on workdays and off-days with a large amount of data, and MSVR is adopted to forecast the KPIs on holidays with few samples. To assess the effectiveness of the proposed method, experiments are conducted in two public datasets with several state-of-the-art methods. The results reveal the superiority of the proposed method in industrial KPIs forecasting.