Advancing neural network architectures for time series forecasting: A sustainable approach to intensified biobutanol production

Abstract

Sustainable production aligning with Industry 4.0 and the UN 2030 agenda is crucial for ecological balance and socio-economic progress. Accurate chemical process forecasting is vital in optimizing operations and cost reduction. We introduce a statistical methodology for determining Long Short Term Memory (LSTM) network architectures, focusing on distillation columns. Using Aspen Plus Dynamics, we generate datasets from a sustainable distillation column separating fermentation-derived effluents into acetone, butanol, and ethanol (ABE) for spark-ignition. Input variables (reflux ratio and reboiler heat duty) and output variables (acetone, butanol, ethanol purities) are analyzed. Results show a one-layer neural network effectively predicts ABE concentrations, with ADAM and RSMP as optimal training algorithms and five neurons to prevent overfitting. Linear activation outperforms hyperbolic tangent functions. Reflux ratio and reboiler duty alone sufficiently capture intensified column dynamics, eliminating the need for three features conventionally. This pioneering methodology demonstrates AI's potential in chemical engineering processes and beyond.