Objective The humus beneath Larix gmelinii forests is rich in organic matter, making it prone to concealed and persistent smoldering fires. As an indicator directly reflecting the dynamic equilibrium between heat generation and dissipation, the heating rate is a key parameter for quantifying heat accumulation characteristics during smoldering. Current systematic research on this parameter remains limited. Furthermore, existing prediction models struggle to balance prediction accuracy with algorithmic convergence efficiency when characterizing the smoldering heating process. This study aims to investigate the response patterns of smoldering heating rates under multi-factor coupling effects. We constructed an improved BO-XGBoost prediction model tailored for predicting smoldering heating rates, providing technical support for forecasting smoldering fire behavior in forest understories.

Method We constructed a dataset of 750 samples based on laboratory smoldering tests of Larix gmelinii humus. Data augmentation was performed by introducing 5% high-intensity Gaussian noise to simulate fluctuations in heating rate data caused by environmental disturbances in wild forest settings. We employed one-way analysis of variance combined with Pearson correlation tests to quantify the independent effects and interactions of moisture content, particle size, and wind speed on heating rates. The core strategy of the model involves introducing Chebyshev mapping for nonlinear transformation to uniformly generate initial optimization sample points, thereby expanding the initial parameter search space. We developed a multi-fidelity optimization strategy that combines low-fidelity rapid search with high-fidelity precise calculation, balancing hyperparameter search efficiency with prediction accuracy. Additionally, a dynamic adaptive weight update mechanism was adopted to dynamically adjust the smoothing coefficient based on weight changes, accelerating Bayesian hyperparameter convergence. We conducted multi-dimensional performance comparisons between the improved model and baseline models (BO-XGBoost, GA-XGBoost, and SSA-XGBoost). Ablation studies were also performed to evaluate the contribution of each module.

Result (1) Single-factor analysis indicated that the heating rate was significantly negatively correlated with moisture content (P < 0.05) and extremely significantly negatively correlated with particle size mesh number (P < 0.01). This suggests that increased moisture content and reduced particle size inhibit smoldering heating in humus. Wind speed exhibited a dual effect on smoldering heating rates: it promoted heating at 0–6 m/s (P < 0.05), but inhibited the process once wind speeds surpassed 6 m/s. (3) Model performance comparisons showed that, compared to the three control models, the improved BO-XGBoost model increased R² by at least 2.14%, reduced MAE by at least 24.54%, increased S² by at least 2.09%, and reduced RMSE by at least 12.08%. This demonstrates higher fitting accuracy and stability. (4) Ablation studies indicated that Chebyshev mapping, the multi-fidelity strategy, and adaptive weights balanced sample distribution, reduced prediction errors, and accelerated model iteration convergence, respectively.

Conclusion The study demonstrates that the smoldering heating rate of Larix gmelinii humus is jointly regulated by moisture content, particle size, wind speed, and their interactions. Conditions characterized by low moisture content, large particle size, and moderate wind speed effectively promote heat accumulation and release in humus, thereby accelerating the smoldering heating process. The improved BO-XGBoost prediction model constructed in this study can simulate the smoldering heating rate of humus under controlled laboratory conditions. It provides a technical solution for optimizing algorithms and improving accuracy in prediction models for smoldering fire behavior in forest understories.