No abstract

Prediction accuracy is crucial for higher heating value (HHV) models to promote renewable biomass energy, especially its consistency is crucial when retraining data and knowledge of the range are unavailable. Current HHV models lack consistency in accuracy and interpretability due to various reasons. Thus, this study aimed to construct an interpretable and consistent proximate-based biomass HHV model on a wide-range dataset. The model, regime-lasso, integrated the concepts of regime-switching, lasso regression, and federated averaging to construct a consistent HHV model. The regime-switching partitioned the dataset into optimal regimes, and the lasso trained the regime models. The regime-lasso model is a collection of these models. It provided root  mean square error of 0.4430– 0.9050, mean absolute error of 0.2743–0.6867, and average absolute error of 1.512–4.5894% in the literature’s wide-range datasets. The Kruskal–Wallis test confirmed the in-sample performance consistency at α=0.05, regardless of the training sets. In the out-of-sample situations without retraining, the model preserved its accuracy in six out of 11 datasets at α = 0.01. The interpretability of regime-lasso indicated the regime characteristic to be a factor of inconsistent prediction. The increase in FC had the maximum positive impact on HHV in the 2nd and 3rd regimes, while the increase in ASH negatively impacted the 1st and 2nd regimes. VM variation had neutral effects in all regimes. The regime-lasso solves the issues of accuracy declination and addresses the challenges in sensitivity analysis of the HHV model. The prediction accuracy issues of the model’s direct implementation were fixed.