Residential biomass combustion is characterized by strong temporal variability, which complicates emission assessment and limits the effectiveness of conventional steady-state models. In this study, time-aware data-driven models were developed to dynamically predict flue-gas composition during operation of a residential woodstove fired with two common biomass fuels: beech wood logs and commercial wood briquettes. Experimental measurements were carried out on a Romotop Lugo N stove under controlled laboratory conditions, with continuous monitoring of operational parameters and gaseous emissions, including CO, CO₂, NOₓ, SO₂, and organic gaseous compounds (OGC). The recorded time series were pre-processed and segmented into distinct combustion phases (ignition, stable combustion, and burnout) to explicitly account for transient behaviour. Several machine-learning approaches were evaluated, including multilayer perceptrons, gradient-boosting models, and hybrid stacked architectures. Model performance was assessed using both stratified validation and chronological validation to reflect real-world forward prediction scenarios. The results demonstrate clear fuel-dependent differences in emission dynamics and model complexity requirements. Briquette combustion exhibited higher stability and could be accurately described using simpler models, whereas beech wood combustion required higher-capacity architectures to capture rapid transitions and non-linear effects. Chronological validation revealed limited generalisation of static multi-output models, highlighting the necessity of time-aware and phase-conditioned modelling strategies. Feature relevance analysis consistently identified flue-gas temperature, oxygen concentration, and combustion phase as dominant predictors across emission species. The proposed framework provides a practical basis for real-time emission prediction and diagnostic support in residential biomass heating systems, contributing to the development of intelligent, low-emission combustion technologies.