Avaliação do modelo LASSO na determinação de metano, etano e propano em gás natural e biogás utilizando um espectrômetro de baixo custo e portátil

Abstract

Natural gas is a fossil fuel, composed mainly of methane, ethane and propane, which is used as a heat and energy source in homes and industries. NG quality control is carried out to determine the methane, ethane and propane contents, in addition to controlling the gas characteristics. For quality control, gas chromatography is commonly used.. However, this technique has some disadvantages such as analysis time, equipment maintenance costs, reagent demand, among others. An alternative for determining these compounds in gaseous mixtures is Near Infrared (NIR) spectroscopy. However, such spectra need chemometric tools so that they can be used for quantitative purposes. Therefore, the Least Absolute Shrinkage and Selection Operator (LASSO) method, which can be used for selection of variables and prediction of future samples, was used to determine methane, ethane and propane in natural gas and biogas samples. The results thus obtained were compared with those obtained by the full spectrum Partial Least Squares (PLS) methods and by the MLR model with the variables selected by the Successive Projection Algorithm (SPA). Data were preprocessed using second derivative with Savitzky-Golay smoothing fitted to 21 point windows and 2nd order polynomial. The parameters obtained in the prediction, rPred and RMSEP, for LASSO were compared with the values obtained for the PLS and SPA-MLR models. For methane, the values obtained by LASSO were similar to PLS and SPA-MLR, with RMSEP of 4,45% mol mol-1 and r of 0,96. For ethane, the LASSO model presented a similar result to PLS and superior to SPA-MLR, with RMSEP values of 3,98% mol mol-1 and r of 0,95. Considering propane, it was observed that LASSO values did not differ significantly from PLS and SPA-MLR, with RMSEP of 1,46% mol mol-1 and r of 0,97. The F test at a 95% confidence level showed that there were no statistically significant differences between the studied methodologies. Therefore, the study achieved its initial objective, which is to demonstrate that LASSO can be used to predict the analytes in a satisfactory way, in the same way as the PLS and S PAMLR models.