Forecasting biomass hydrogen production using a novel recency-weighted grey multivariate model

The scalable production of green hydrogen from biomass gasification is hindered by unpredictable yields arising from complex, non-linear interactions among multiple feedstock and process parameters, compounded by sparse data in pilot-scale operations. Existing machine learning models demand large datasets, while traditional grey system models lack adaptability to recent information and multivariate dependencies. This study introduces a novel recency-weighted grey multivariate model incorporating a data preference accumulation operation with exponential weighting controlled by memory factor to dynamically prioritize recent observations. Integrated into a multivariable grey framework, it captures interdependencies while maintaining robustness with small samples. Theoretical stability analysis shows enhanced noise resistance in limited-data regimes. Validated on real-world biomass gasification data with 23 data points, the novel model achieves high accuracy with 0.096% mean absolute percentage error, 0.046 root mean square error and coefficient of determination of 0.9907. The prediction robustness is confirmed by Diebold-Mariano and superior prediction ability tests over some grey models, statistical models, and machine learning benchmarks. This first application of recency-weighted grey multivariate forecasting to biomass hydrogen production provides a data-efficient, interpretable tool for optimizing reactor operation and feedstock selection, accelerating the transition to a sustainable hydrogen economy.