Data driven models for online tar monitoring in biomass gasification

Reliable real-time monitoring of tar concentration remains a major barrier to process control and downstream syngas utilization. While online tar measurement is often constrained by high system complexity and associated costs, data-driven soft sensors offer a feasible and efficient alternative. Existing approaches often treat tar concentration in its raw form, despite its highly skewed distribution, which can compromise model stability, generalization, and the physical reliability of soft-sensor predictions. This study proposes a data-driven soft-sensor pipeline for the online estimation of tar concentration, combining domain-specific feature engineering, target-variable transformation, and deployment-oriented latency assessment across 11 ML algorithms, one DL model, and a Stacking Regressor ensemble, using a dataset comprising 221 literature-derived experimental records. The proposed pipeline systematically assesses the impact of four target variable transformations (logarithmic, Box-Cox, Yeo-Johnson, and untransformed raw data). Results demonstrate that target variable transformation is crucial for model stabilization and generalization. On the independent test set, the Stacking Regressor model coupled gradient and categorical boosting regressors achieves the best predictive performance (R 2 = 0.85, RMSE = 4.06 g/Nm 3 ). Similarly, the single categorical boosting regressor coupled with Box-Cox transformation proves to be a robust alternative (R 2 = 0.83). In addition, all top-performing models met the 1 Hz real-time monitoring constraint, indicating the feasibility of the proposed pipeline for soft-sensor deployment in biomass gasification systems. Overall, the findings show that boosting and ensemble nonlinear models combined with appropriate target-variable transformation offer a better balance between accuracy, generalization, and computational feasibility than more complex architectures, highlighting their potential for use as soft sensors for real-time tar monitoring in biomass gasification systems.