Diseño de una red neuronal para la predicción de la producción de metanol en una planta de hidrogenación de CO2.

Abstract

The objective of this research work was to develop an artificial neural network with the ability to predict with reasonable precision the production of methanol that would be obtained from a CO2 hydrogenation plant. For the construction of the network, it was first modeled through the multiplatform chemical process simulator compatible with windows, DWSIM, a CO2 hydrogenation plant based on that developed by Éverton Simões Van-Dal and Chakib Bouallou, once the plant was simulated. From this 100 simulation data, varying the flow of the reagents as well as their temperatures and pressures through a combination of random numbers, statistical tests were applied to these 100 data to verify the non-existence of atypical data. In the Matlab software, these previously normalized data are entered and with the help of the Neural Network Fitting tool, the neural network is developed and, through a trial and error process, the optimal architecture of the network is determined based on the minimum mean square error. Finally, a statistical comparison was made between the data provided by the simulator and the network output values to determine significant differences as a validation method. As a result, a network of 12 hidden neurons, 4 input neurons and 1 output neuron was obtained with a mean square error of 0.008506 whose predicted values do not differ significantly with the values provided by the simulator. It is concluded that the designed network is capable of predicting the methanol flow that would be obtained from a hydrogenation plant with the same accuracy as a simulator would. It is recommended to add the separation pressure in the recirculation zone as an input parameter as a way to obtain a result that is closer to reality.