The precipitation of gas hydrate and inorganic salts (scale) during oil and gas production represents a significant flow assurance hindrance for the industry. Chemical inhibitors can prevent the fouling process, but specific inhibitors to address a problem could result in synergistic or adverse effects. Simulations in tubes and pipelines are necessary to understand these behaviors by assessing the scaling tendency of the water. The primary objective of this study was to create models using an artificial neural network (ANN) of the multilayer perceptron (MLP) type for the simulation of the calcium carbonate scaling formation process in the presence of monoethylene glycol (MEG), a typical gas hydrate inhibitor. A database was obtained from 38 tube blocking test (TBT) experiments with different conditions. The models were developed using MATLAB R2020a, splitting the database into two groups on the ratio of 70:30, respectively, train and test ones, preserving the time dependency of the differential pressure (ΔP) data. The ANNs were created using six inputs (temperature, pressure, calcium and bicarbonate concentrations, MEG concentration, and ΔP measured at a selected time) and one output (ΔP measured at a later time). The goal was to explore how monitoring the conditions in a pipeline can predict the evolution of the scaling process. We investigated two scenarios for the ΔP prediction: a near future (one step ahead) and a far future (five steps ahead). The MLP models demonstrated high performance, with an R 2 higher than 92.9% for both training and test groups for both prediction horizons. Then, these models were tested with a second data group to evaluate their applicability to control the systems. The best models showed good scaling prediction, with R 2 ranging from 80.0% to 99.9%. These results represent a promising step toward applying machine learning techniques to simulate and predict scaling tendencies in controlled pipelines.
Modelling a process or equipment is a profitable strategy to build better control strategies, predict fault conditions, and optimize the processes. Different approaches could be explored to achieve the development of better models. This paper investigates the use of experimental data generated by a central composite rotatable design (CCRD) to develop models capable of predicting the performance of a three‐product hydrocyclone for several setups with different dimensional parameters values. Two different modelling strategies are explored: response surface methodology (RSM) and artificial neural networks (ANN). With the RSM models, it was possible to evaluate the statistical importance of the input variables to each output variable. The ANN models showed improved coefficients of determinations (R2) compared to the RSM models, presenting values higher than 97% for all cases, while the RSM models ranged from 79.07%–88.83%. The ANN was demonstrated to be the most effective method to model the physical problem of three‐product hydrocyclones, and it captured its non‐linearities. It was shown that the combination of the design of experiments and ANN to analyze this physical problem is successful and may also be applied to other problems. As far as we have knowledge, a work regarding the comparison of both RSM and ANN methods applied to three‐product hydrocyclones was not found in the literature; this absence was the motivation for this work.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.