A probabilistic deep learning approach for thermal and exergy forecasting in organic Rankine cycles

Abstract

Energy recovery from waste-energies sources is one of the main approaches for reducing carbon footprint and enhance sustainability indicators. One of the main challenges faced when dealing with complex processing systems refers to the development of reliable process models. Such a models are required for simulation, optimization, design and control tasks. When those models are time consuming to develop, or the basic underlying behaviour is not well understood, the leverage of surrogate models is a feasible alternative for addressing the past tasks. For surrogate modelling aims, we deploy feedforward fully connected deep learning nonlinear mappings, to learn process behaviour to forecast thermal and exergy efficiencies from a set of noisy measurements. We also show that common feedforward deep learning models are not reliable for extrapolation aims, and make some proposals to deal with this issue. Overall, this work aims to boost process intensification tasks in the energy recovery from exhaust or waste-energy sources commonly available in processing plants.