Aprendizado de máquina aplicado à analise da operação de uma planta piloto para gaseificação de carvão mineral

Abstract

The gasification process consists in the reaction between mineral coal and one gasifying agent to generate an energetic gas, called synthesis gas (syngas). The reactor where this process occurs is called gasifier. The gasification process can be studied based in operational conditions. The input conditions are: air flow rate in the different regions of the reactor, coal flow, heating element operation, manual pressure pulses and centrifuge frequency. The output conditions in the process are the temperatures and the pressures obtained through sensors, which are located in the course of the reactor bed. In this case study machine learning techniques are applied to create algorithms capable of making predictions through models based on data. This field has been standing out because of its potential in analisys of industrial processes. The aim of this work is to evaluate the aplication of different machine learning algorithms to predict the operation of a fluidized bed gasifier with bubbling air, which is intalled in the Laboratório de Energia e Carboquímica of UNIPAMPA, Campus Bagé, using mineral coal from Candiota – RS. In order to develop a solid database 15 experimental runings were selected from the experiments records. The data since 2016 until 2020 were collected. The acquisition of the output data was executed through a computer placed in the pilot plant. The input data were obtained from a logbook with laboratory record. To the computational analisys, the development enviroment chosen was Jupyter Notebook, where Python programming language was employed. To implement the current work Python libraries were applied, Pandas for manipulation and analisys of the data, Numpy for manipulation of structures and numerical data and Matplotlib for data visualization. The libraries Scikit-learn and FireTS were used to employ machine learning algorithms. Using these tools the data was preprocessed, the correlation between the variables of operation was analysed and the following algorithms were applied: decision tree, random forests and non-linear autoregressive neural network with exogenous inputs (NARX). Most of the variables in the process showed high correlations between each other, it can be explained by physical and chemical process characteristhics. The machine learning performed partially satisfactory to the decision trees and random forests models, since were obtained R² values of 0.8263 and 0.8262 respectively. However the graphical analisys showed that the predictions doesn’t match the dynamics of the real process. To the NARX model, the result was satisfactory, getting to R² values of 0.9955 to temperature predictions, and 0.8789 to pressure predictions. The gasification process is extremely dinamic, for this reason the neural networks of the NARX model present better results, since it carries memory and is trained to learn from the patterns of time variation.