Development of Neural Model of Gas Processing Plant

Resumo

The total volume of gas processed in Brazil is forecast to increase over the next decade, according to the ten-year energy expansion plan (EPE). As a result, the occurrence of failures in gas processing could result in the processed gas being out of specification and/or lead to the interruption of gas processing operations (availability and reliability), especially in a scenario where the gas comes from different extraction points, with different characteristics.  Given the complexity of the plant and the existing non-linearities, the use of Deep Neural Networks (Deep Learning) was sought to build a neural model of a gas processing plant, to predict dynamic behavior, with the prospect of using the evaluated neural model for continuous monitoring of operations (Digital Twin). The choice of neural networks follows a more current and fast-response approach, more suitable for continuous monitoring, when compared to the use of rigorous phenomenological models, with long response times. A real dataset, obtained from data collected from a gas processing plant (UPGN), will be used, with the generation of neural models for the prediction of dynamic behavior. The performance of the various neural models will be compared. The models will be tested across multiple prediction horizons to determine the effectiveness of multi-step-ahead predictions.