C-Cast: A Real-Time Forecasting Model for a Controlled Sequence

Abstract

Predictive control is an advanced control method that is used successfully in industrial control applications. One of the most fundamental demands for predictive control is the accurate forecasting of a ''controlled sequence" using exogenous sequences which consist of multiple attributes (manipulated variables and operation signals). Given a controlled sequence and exogenous sequences, how can we effectively forecast the future behavior of a controlled sequence? In this paper, we present C-Cast, an efficient and effective method for forecasting a time-evolving controlled sequence with exogenous sequences. Our proposed method has the following properties: (a) Adaptive: it captures important time-evolving patterns and operation shift in a time-evolving controlled sequence (b) Effective: it performs accurate forecasting. (c) Practical: it enables real-time controlled sequence forecasting fast enough to satisfy the limitation required for predictive control. Extensive experiments on a real dataset demonstrate that C-Cast consistently outperforms the best existing state-of-the-art methods as regards accuracy, and the execution speed is sufficiently fast.