Thermal runaway of ethylene oxidation reactors: Prevision through neuronal networks

Abstract The dynamic behavior of an ethylene oxidation fixed-bed reactor has been originally simulated by a phenomenological model, encompassing mass and energy balances of the catalytic bed. This model makes use of the one-dimensional pseudo-homogeneous approach, with apparent kinetic parameters obtained from the literature. The resulting set of partial-differential equations is solved by discretization of the space variable in finite-differences and integration of the attained ordinary-differential equations with respect to time with a marching algorithm that accounts for the problem of stiffness near the runaway point. This paper focuses on the use of a neuronal network in forecasting possible thermal runaway situations of this highly exothermic process. The final objective is to build a reliable inference alarm algorithm for fast detection and prevention of this situation. The neuronal network also predicts eventual hot spot positions and temperature, based on informations such as inlet flows, temperatures and pressures, provided by the plant instrumentation. Feedforward neuronal networks were used, with one hidden layer. A training algorithm based on a combination of backpropagation and gaussian random guesses was applied. The neuronal network represents well the evolution of the transient temperature profile.