(635e) Machine Learning-Based Soft Sensors for Vacuum Distillation Unit

Traditional approaches for refinery process optimization are usually based on so-called first-principles models. First-principles models have a number of disadvantages, such as the need for specialized knowledge of the process (which may lead to biased results), complexity imposed by the parameters of models (e.g. kinetics, and mass and heat transfer), the time required to obtain the optimization results and the difficulty in converging these highly non-linear optimization problems. Hybrid models – a combination of first-principles and empirical or statistical models, have been proposed as a solution for the aforementioned issues. Ochoa-Estopier et al. (2015) optimized heat-integrated crude oil distillation systems (CDU) using artificial neural networks (ANN) and reported significant improvement in the total profit [1].

Whilst operational constraints related to refinery units can be enforced using computational techniques during process optimization, the product quality – another significant target – is still often analyzed experimentally, rather than using online analyzers. Some of the problems around assessing the product quality are time delays related to sample withdrawal and time-consuming laboratory measurements. Soft sensors, which predict and forecast product quality based on sensor measurements of properties, such as temperature, pressure and flow rate, have been shown to be good alternatives to laboratory measurements for the assessment of product quality. Shang et al. (2014) developed data-driven soft sensors based on stacked restricted Boltzmann machines for CDU with excellent predictability (deviation of as low as 3^oC for ASTM 95% cut point temperature of heavy diesel) [2].

Machine learning methods (such as ANN) are classed as black box models, i.e., mapping the input-output data with its mathematical foundations not based on knowledge of the underlying physicochemical phenomena. Nevertheless, such models provide speed and performance that cannot be matched by first-principles models. However, the source of data used to establish those models is crucial, e.g. either simulated or real data. Ahmad et al. (2020) reviewed different models and industrial applications of soft sensors. The authors concluded that the pitfalls of black box models can only be avoided by using appropriate frameworks, systematically improved and adapted, including data pre-treatment, feature engineering, selection and importance analyses, model insights or data visualization [3].

This collaborative academia-industry project focuses on developing a framework for real-time optimization of chemical process units via soft sensors using cutting-edge data science techniques. Historical data for a vacuum distillation unit in a Chinese petroleum refinery is used as a case study to illustrate the framework.

In this paper, we employ a variety of methods to address the ways of tailoring machine learning models that originally would be classed as black box, including data pre-treatment, input-output correlations for models explainability.

Methods

The first category of data is plant measurements of temperature (T), pressure (p) and flow rate (F) that make up the input data for soft sensors. The second category is ASTM-D2887 laboratory measurements of the boiling curve of a distillation product of 3^rd side draw vacuum distillation product, namely V3SS; the soft sensors developed use measured data to infer the boiling curve (2, 10, 30, 50, 70, 99 vol%). In this work, we propose a framework for development of robust soft sensors, ensuring reliability, reproducibility and explainability of model predictions.

The first step involves data pre-treatment, more specifically enhancing the quality of the data by identifying and removing outliers and apparent malfunctioning of sensors. Plant measurements are analyzed using the maximum log-likelihood approach combined with principal component analysis, PCA, and k-means clustering to identify short- and long-term outliers and detect readings from malfunctioning sensors. Laboratory measurements are analyzed for predictability threshold/accuracy (application of autoregressive integrated moving average models, ARIMA), outlier detection (statistical inference - boxplot), filling missing values and increasing data frequency (ARIMA, kNN regression).

We use ANN models based on sequential multilayer perceptron architectures. Numerous approaches to improve the models, including the ANN architecture (number of hidden layers, number of nodes in hidden layers) and other hyperparameters (activation functions, learning rates, optimizers, dropout, and regularization layers), were studied in detail. The aim of this comprehensive investigation is to gain insight into the capabilities of basic neural networks for the task and showcase the value of tailoring neural networks for a specified purpose.

The aim of the ANN was to achieve the lowest threshold from ARIMA modelling, considering networks with different architectures. In addition to achieving the best possible result with ANN against the ARIMA threshold, a thorough study was carried out for machine learning models’ explainability. Firstly, PCA was used to reduce the dimensions of the input variables of the neural network (namely feature pre-selection). Feature importance analysis was used to understand how the sensors measurements are correlated in the ANN to reflect the underlying physicochemical behavior of the distillation process. The Shapley Additive Explanations (SHAP) method was used to provide an insight into the feature importance, and further post-selection of features, in order to maximize the efficiency of the network, while minimizing the computational cost caused by the data dimensionality, as well as so-called response surface approach.

Result

Firstly, ARIMA modelling was used to assess the uncertainty of the V3SS product boiling curve (Table A). Those values were used as a threshold to assess optimized ANN performance. We found that the end point (V3SS 100 vol%) is associated with the highest uncertainty (i.e., 9.0°C).

Secondly, the architecture and hyperparameters were shown to have a significant impact on the performance of the networks (Figure A), highlighting the importance of tailoring the ANNs, e.g. mean absolute error (MAE) in the best case using raw data (without any data pre-treatment) is 9.5^oC (for 2 hidden layers x 30 nodes in hidden layers). Changes in learning rate for the optimizers (from 0.001 to 0.01) can slightly improve the performance, i.e. decrease of MAE to 9.3 ^oC (for 3 hidden layers x 30 nodes in hidden layers). Moreover, the increase of learning rate can also improve the training speed; however, a more complex model was required to gain this improvement (i.e. 3 rather than 2 hidden layers). Networks with the same architecture were compared against using pre-treated data. The best performance in terms of MAE was 8.5^oC (for 2 hidden layers x 20 nodes in hidden layers). This improvement by using pre-treated data highlights numerous advantages, including decreased complexity of the network (with pre-treated data, the simpler model performed better ), improved training speed (through adjusted learning rate) and overall better performance in the prediction (1^oC decrease in the mean absolute error of loss function) which makes the network perform to the set threshold (9.0 ^oC from ARIMA). More details on the data pre-treatment techniques and results are given in the presentation.

Thirdly, feature importance was used to retrieve correlations between input and output, giving an understanding of the models’ parameters against the underlying physicochemical phenomena of distillation process. The stream of interest is associated with the 3^rd draw inlet temperature (measured by temperature sensor T3) which, as expected, is highly correlated with the predicted boiling curve cut points for V3SS product (as seen in Figure B, for the most accurate ANN model using pre-treated data). Other vital sensors for the vacuum column also exhibited high importance in the trained models, e.g., bottom and top pressures (PT and PB, respectively) and the volumetric flow rate of the product (F33). Other important results were that:

• Feature pre-selection with PCA leads to preserving the performance whilst reducing the dimensionality of the data.
• Raw data can lead to unreasonable feature importance representation (contrary to the underlying physical understanding of the system).

Implications

Real-time optimization with self-calibration and developing soft sensors are currently of significant interest in the chemical process industries. The application of computational techniques, especially data science, alongside historical plant data, enables meaningful analysis and filtering of the plant data; this step is essential if the data are to be used for process optimization. There is a plethora of literature developing methods for each of these steps, however, little attention has been paid to integrating them and comparing their efficacy. In this work, we addressed some of the issues that one can encounter when developing models based on machine learning approaches, such as:

• The impact of data pre-treatment on development and performance of soft sensors
• How to set the threshold for the models’ performance and accuracy
• How to achieve set performance thresholds through optimization of ANN structures and appropriate data pre-treatment
• How to avoid ‘black box’ models through models explainability which, ideally, is well-aligned with the process physical and chemical understanding.

References

[1] Ochoa-Estopier and Jobson, Industrial and Engineering Chemistry Research, 54, 2016, https://pubs.acs.org/doi/10.1021/ie503802j

[2] Shang, Yang, Huang and Lyu, Journal of Process Control, 24, 2014, https://dx.doi.org/10.1016/j.jprocont.2014.01.012

[3] Ahmad, Ayub, Kano and Cheema, Processes, 8, 2020, https://doi.org/10.3390/pr8020243