Multiple target data-driven models to enable sustainable process manufacturing: An industrial bioprocess case study

Abstract

Abstract Process manufacturing industries constantly strive to make their processes increasingly sustainable from an environmental and economic perspective. A manufacturing system model is a powerful tool to holistically evaluate various manufacturing configurations to determine the most sustainable one. Previously models of process manufacturing systems are typically single target models, trained to fit and/or predict data for a single output variable. However, process manufacturing systems produce a variety of outputs with multiple, sometimes contradictory, sustainability implications. These systems require multiple target models to find the most sustainable manufacturing configuration which considers all outputs. A novel bioprocess that treats process wastewaters to reduce pollutant load for reuse, while simultaneously generating energy in the form of biogas was studied. Multiple target models were developed to predict the percentage removal of chemical oxygen demand and total suspended solids, in addition to the biogas (as volume of methane) produced. Predictions from the models were able to reduce wastewater treatment costs by 17.0%. Eight models were developed and statistically evaluated by the coefficient of determination (R2), normalised root mean square error (nRMSE) and mean absolute percentage error (MAPE). An artificial neural network model built following the ensemble of regressor chains demonstrated the best multi target model performance, averaged across all the bioprocess’s outputs (R2 of 0.99, nRMSE of 0.02, MAPE of 1.74). The model is able to react to new regulations and legislation and/or variations in company, sector, world circumstances to provide the most up to date sustainable manufacturing configuration.