Steam consumption prediction of a gas sweetening process with methyldiethanolamine solvent using machine learning approaches

Abstract

This study proposes a comprehensive data processing and modeling framework for building high-accuracy machine learning model to predict the steam consumption of a gas sweetening process. The data pipeline processes raw historical data of this application and identifies the minimum number of modeling variables required for this prediction in order to ease the applicability and practicality of such methods in the industrial units. On the modeling end, an empirical comparison of most of the state-of-the-arts regression algorithms was run in order to find the best fit to this specific case study. The ultimate goal is to leverage this model to identify the achievable energy conservation opportunity in such plants. The historical data for this modeling was collected from a gas treating plant at South Pars Gas Complex for 3 years from 2017 to 2019. This data gets passed through a multistage data processing scheme that conducts multicollinearity analysis and model-based feature selection. For model selection, a wide range of regression algorithms from different classes of regressor have been considered. Among all these methods, the Gradient Boosting Machines model outperformed the others and achieved the lowest crossvalidation error. The results show that this model can predict the steam consumption values with 98% R-squared accuracy on the holdout test set. Furthermore, the offline analysis demonstrates that there is a potential of 2% energy saving, equivalent to 24 000 metric tons of annual steam consumption reduction, which can be achieved by mapping the underperforming energy consumption states of the unit to the expected performances predicted by the model.

A machine learning‐based operational control framework for reducing energy consumption of an amine‐based gas sweetening process

“…For this purpose, the predictive models have been developed based on an ML-based modeling framework proposed in our previous study. 15 In that work, the GBM regressor model has been presented as the most suitable algorithm for this use case that meets the high-precision forecasting ability required for this application. The dataset containing input and manipulated variables was used to develop the predictive models.…”

“…The minimum independent variables for ML modeling have been determined based on a data processing pipeline described in our previous study. 15 The input variables are the most important predictive features related to the steam consumption of the GSP, and have been selected from absorption column and input steam condition. Input variables can affect the regeneration process and change the mode of the process.…”

“…The robust and highly accurate prediction models are needed in order to estimate the controlled and output variables and to develop a reliably efficient control strategy. For this purpose, the predictive models have been developed based on an ML‐based modeling framework proposed in our previous study 15 . In that work, the GBM regressor model has been presented as the most suitable algorithm for this use case that meets the high‐precision forecasting ability required for this application.…”

Power prediction of waste heat recovery system for a cement plant using back propagation neural network and its thermodynamic modeling

“…For each algorithm, a hyper-tuning parameter was run for optimizing the results, and a gradient boosting machine regressor showed the most promising result in terms of the least mean absolute percentage error (MAPE) among all techniques followed. 26 Similarly, results from different neural networks were compared by Kontoggiannis et al for predicting the power consumption from a residential dataset. 27 Multi-layer perceptron was found to be the better option, as it converged the data fast and gave fairly accurate results of prediction.…”

“…To overcome different practical challenges, 10 different machine learning approaches were compared for predicting the steam consumption of a gas‐treating plant. For each algorithm, a hyper‐tuning parameter was run for optimizing the results, and a gradient boosting machine regressor showed the most promising result in terms of the least mean absolute percentage error (MAPE) among all techniques followed 26 . Similarly, results from different neural networks were compared by Kontoggiannis et al for predicting the power consumption from a residential dataset 27 .…”