Neural Computing and Applications (09410643)35(11)pp. 8517-8541
Using a reliable predictive model is important for modeling, controlling, and optimization of the isomerization process. This process has a significant impact on the gasoline quality, which can reduce greenhouse gases by improving the octane number. On the other hand, the accuracy of the predicted results of a data-driven model depends on the quality of input data; this is while the measured variables of industrial units are inevitably contaminated by various errors. Hence, the present work proposes an improved adaptive machine learning model and a new hybrid multiscale filter to predict the gasoline research octane number reliably from error-contaminated data of a light naphtha isomerization reactor. The proposed machine learning model is based on the integration of the feature selection algorithm of the double-level similarity with the support vector regression model (named DLS-SVR model) for adaptive prediction. The new hybrid filter is based on a combination of the wavelet transform and median absolute deviation, named multiscale median absolute deviation (MSMAD). MSMAD filter is proposed with the aim to establish an accurate method to identify and eliminate outliers and gross errors from the measured process variables. A pilot-scale reactor is employed to provide the required experimental validating dataset to evaluate the predictive performance of the proposed filter–model combination. Inputs of the DLS-SVR model are operating conditions (temperature: 115–150 °C, pressure: 28–42 bar, space velocity: 0.38–3 h−1) and feed composition (benzene: 0–3.5 wt%, cyclohexane: 0.8–23.2 wt%, methylcyclopentane: 1–29 wt%, H2/naphtha ratio: 0.03–0.3). The performance of the DLS-SVR model is compared with the response surface methodology, support vector regression, and double-level locally weighted extreme learning machine through the fivefold cross-validation technique. The particle swarm optimization–sequential quadratic programming algorithm is used to optimize the hyper-parameters of these models. The results prove that the generalized DLS-SVR model outperforms the other generalized models. Furthermore, the performance of the MSMAD filter is compared with the multiscale median, finite impulse response–median hybrid, median, and median absolute deviation filters by rectifying the error-contaminated temperature signal. Findings reveal that the DLS-SVR model utilizing the rectified signal by the MSMAD filter has a maximum coefficient of determination, R2 = 0.91, and minimum root mean square error, RMSE = 0.0562, among the other filter's rectified temperature signals. These values for error-free data are R2 = 0.945 and RMSE = 0.0439. © 2022, The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature.
