Deep learning for industrial processes: Forecasting amine emissions from a carbon capture plant

Abstract

One of the main environmental impacts of amine-based carbon capture processes is the emission of the solvent and degradation products into the atmosphere. To mimic the mounting importance of intermittent operations of power plants we performed a stress test in which we measured the amine emissions from a pilot plant that has been in operation using a mixture of amines (CESAR1) in a slipstream from a coal-fired power plant. Understanding how changes in the operation far from the steady-state of the plant affect the emissions is key to designing emission mitigation strategies. However, conventional process modelling techniques struggle to capture the full dynamic, multivariate, and non-linear nature of this data. In this work, we report how a data-intensive approach can be used to learn the mapping between process and emissions from data. The resulting model can forecast the emissions, can be used to analyse the data and also perform in silico stress tests. By doing so, we reveal that emission mitigation strategies that work well for single component solvents (e.g. monoethanolamine) need to be revised for a mixture of solvents such as CESAR1. We expect that the combination of large amounts of data with flexible learning algorithms will impact the way we design and operate industrial processes, as we can now harvest information at conditions where conventional approaches fail.