Hybrid Data-Driven and Mechanistic CO2 Soft Sensor with MHE-Imputed Labels and Covariance-Weighted Fusion in a Pilot-Scale Absorber

Gas analyzers in post-combustion CO2 capture plants are accurate but slow and sequential, yielding sparse, non-synchronous concentration records across absorber stages. We address this missing-data problem by reconstructing continuous CO2 profiles with Moving Horizon Estimation (MHE) constrained by a mechanistic absorber model and available measurements; these MHE reconstructions are used as supervisory labels to train an end-to-end Stacked Denoising Autoencoder–Gated Recurrent Unit (SDAE-GRU) model. At run time, we deploy a hybrid soft sensor using the SDAE-GRU together with the mechanistic model and fuse their open-loop predictions via covariance-weighted blending with Gaspari-Cohn localization. We validate this approach on a pilot-scale MEA absorber using data from seven pilot runs conducted at distinct operating conditions, using datasets 1–5 for training/tuning and 6–7 for blind validation. On the blind validation runs, the hybrid estimator achieves a MAPE of 3.79% for stage-wise CO2 predictions (averaged over all stages and time samples), outperforming both constituents evaluated standalone: 7.86% for the GRU-only soft sensor and 6.79% for the mechanistic model. Because MHE is used only offline to generate labels and to estimate model-error covariances, the deployed estimator is lightweight and suitable for online monitoring.