Neuro probabilistic model for the prediction of the air demand in N2-PSA plants including the quantification of its uncertainty

N 2 -Pressure Swing Adsorption (PSA) is a process that efficiently generates pure N 2 from compressed air. However, its design and control strategy are non-trivial due to multivariable interactions between ambient condition, demand situation, and different process variables. Therefore, it is reasonable to expect potential benefits from machine learning tools. As the compressor is the dominant power consumer in these plants, this study presents a neuro probabilistic model for the prediction of the air demand and the quantification of its uncertainty. The authors demonstrate that data point estimators, such as the Gradient Boosting Regressor (GBR), the Random Forest Regressor (RFR) and the Multi-Layer Perceptron (MLP), trained under the assumption of homoscedasticity, can perform well according to the statistical evaluation metrics. However, by analysing the scatter plot and the residuals distribution, the process's heteroscedasticity is evident. Consequently, deterministic predictions do not provide rigorous uncertainty quantification, leading to decreased reliability. Conversely, the Probabilistic Neural Network (PNN) we propose is a valuable approach that provides a distribution of possible outcomes, enabling the prediction of the air demand and the quantification of the aleatoric uncertainty.