Time series forecasting for dynamic scheduling of manufacturing processes

Manufacturing control systems evolved in the recent decades from pre-programmed rigid systems to adaptable, data driven, cloud based implementations, capable to respond to environment changes and new requirements in real time. A byproduct of this transformation is represented by large amounts of structured and semi-structured information, both historical and real-time data that is made available on various layers of the system. This accumulation of information brings the opportunity to move from the rule based decision making algorithms used traditionally by these control systems towards more intelligent approaches, driven by modern deep learning mechanisms. This paper proposes a time series forecasting model using recursive neural networks (RNN) for operation scheduling and sequencing in a virtual shop floor environment. The time series aspect of the RNN is novel in manufacturing domain, in the sense that the new best prediction produced considers the previous decisions and outcomes. The proposed implementation explains how the RNN can be mapped to the specifics of a manufacturing control system and introduces a bidding mechanism to allow dynamic evaluation of individual forecasts. The pilot implementation, initial experiments on sample data sets and results presented show how using recursive neural networks can optimize resource utilization and energy consumption.