A Neural Network Approach for Prediction of Main Product Yields in Methanol to Olefins Process

Abstract

The transformation of methanol into olefins (MTO) is gaining interest in view of the strong demand for light olefins by the petrochemical industry. Modeling of MTO process is required for both reactor design and process control. In this paper, a three layer perceptron neural network model is used to predict main product yields in methanol to olefins conversion. The network is fed with three inputs, i.e. the reactor temperature, water/methanol ratios in feed, and space-time to predict the weight fraction of unconverted methanol, ethylene, propylene and butanes. The optimum structure of neural network (NN) is determined by a trial and error method. The performance of basic backpropagation (BBP) training algorithm is compared with backpropagation with declining learning-rate factor algorithm (BDLRF). It is found that BDLRF has a better performance in both training and test phases. The network with optimum topology learns the input-output mappings with enough accuracy for interpolation cases. By using NN model in MTO process developing complex reaction kinetic in both global and hybrid model can be avoided.