Introducing the state space bounded derivative network for commercial transition control

The state space bounded derivative network (SSBDN) represents a significant evolution in universal approximating (UA) technology that has important implications for the future automation of polymer grade transitions. Until now, existing UA technologies such as neural networks have had rudimentary architectural deficiencies that result in highly inaccurate and unusable process gain predictions over dynamic transitions (Turner 2002). In such circumstances, the neural element is switched off and laboratory feedback is used to derive the process to its final setpoint. This considerable inefficiency (adding many hours to transition times) has been addressed and solved by the SSBDN technology which provides a powerful, universal approximating technology that can accurately and robustly predict process gains and dynamics over grade transitions resulting in truly predictive models that drive the process to setpoint by anticipating rather than reacting to laboratory feedback. In addition to this, because of its intrinsically safe architecture, the SSBDN can be placed within the kernel of the steady state and dynamic optimisation. This results in a full nonlinear optimisation at each control cycle. This paper will introduce this exciting new technology and present details of its successful implementation on a commercial polyethylene facility which resulted in record breaking transition times.