The influence of the activation function of a neural network on the approximation of data from the main control channels during the start-up of an acetic acid synthesis reactor

Abstract

Modern industrial technologies for the synthesis of acetic acid and its derivatives are mainly based on highly efficient catalytic methods, among which methanol carbonylation (Monsanto and Cativa processes) occupies a leading position. The efficiency of such production facilities directly depends on the operation of chemical reactors, which form the hardware basis of the industry. In a highly competitive market, it is critical to ensure strict compliance of products with industry specifications, which requires accurate monitoring and control of the input and output parameters of the technological process.

 The most common approach involves using neural networks as black box models. Based on empirical data from the object, the network architecture is selected and trained (weight coefficients are adjusted) to achieve maximum model adequacy. The models obtained will subsequently be integrated into predictive control systems (Model Predictive Control, MPC).

The start-up of an acetic acid synthesis reactor (in particular, in the process of methanol carbonylation on a rhodium catalyst) to operating mode is a complex dynamic process. During start-up, the system transitions from an inert state to an active reaction phase, accompanied by a sharp increase in pressure and an intense exothermic effect.

The use of artificial neural networks at this stage makes it possible to minimise overregulation — a critical control problem that threatens to trigger emergency protection systems and reset parameters.

Numerical modeling was performed in MATLAB (version 2021) using an iterative procedure for structural synthesis of a neural network. The empirical basis for the study was provided by statistical data on the dynamics of the acetic acid synthesis reactor during the start-up period.

Modeling errors for carbon monoxide are almost identical to the results for methanol, indicating the symmetry of the effect of both reagents on the reactor output parameters. For pressure, temperature, and level control channels, regardless of the input influence (methanol or CO), the optimal architecture is one with a linear purelin activation function on the output layer. This ensures high approximation accuracy with an error of no more than 1.6%.

The acetic acid concentration channel proved to be the most difficult to model for both inputs (minimum error 53.71%). This indicates that it is not enough to predict product quality at the moment of launching a standard network structure—it is necessary to take into account dynamic delay (inertia) or increase the depth of the network.