Against the backdrop of climate change, forest fires increasingly undermine ecosystem stability and reshape species distributions in Heilongjiang Province. Therefore, quantifying the drivers of fire occurrence and conducting long-term fire risk forecasting holds critical value for regional ecological security. Centered on the forested regions of Heilongjiang Province, this study systematically assessed the relative contributions of multi-source factors—including topography, vegetation, and meteorological conditions—to fire occurrence and compared the predictive performance of three models: Deep Neural Network with Residual Connections (ResDNN), Artificial Neural Network (ANN), and Support Vector Machine (SVM). Modeling results based on historical fire records indicated that the ResDNN model achieved the highest accuracy (85.6%). Owing to its robust nonlinear mapping capability, it performed better in capturing complex feature interactions than ANN and SVM. These results demonstrate its strong applicability to forest fire prediction in Heilongjiang Province. Building on these findings, the study employed the best-performing ResDNN model in conjunction with CMIP6 multi-model climate projections to simulate and map the spatiotemporal probability of forest fire occurrence from 2030 to 2070. The results provide an intuitive representation of long-term fire-risk trajectories under future climate scenarios and offer scientific support for regional fire prevention, monitoring, early-warning systems, and forest management under climate change.

Key message: Integrating ecological process knowledge into machine learning substantially improves the robustness and extrapolation of forest carbon stock predictions. The knowledge-guided model maintained strong predictive performance over 10–15-year horizons, when conventional machine-learning performance declined markedly, providing a promising framework for long-term forest carbon assessment under climate change. Context: Accurate quantification of forest carbon stocks and their dynamics is essential for understanding ecosystem carbon sequestration potential. However, traditional machine learning models often exhibit limited extrapolation capability in long-term forest carbon predictions because they rely primarily on statistical relationships learned from historical observations and lack explicit representation of ecological processes. Aims: This study aimed to develop a knowledge-guided machine learning (KGML) framework that integrates ecological process knowledge with data-driven algorithms to improve the accuracy, stability, and long-term generalization of forest carbon stock predictions. Methods: The study focused on natural forests in Heilongjiang Province, China, defined as stands classified as having a natural origin in forest inventory records. Forest fixed plot observations from 1976 to 2015 were combined with historical and projected climate data and atmospheric CO ₂ concentration data. The CO ₂ fertilization effect was incorporated into an optimized forest carbon sequestration (FCS) process model, and its outputs and process-based constraints were subsequently integrated with machine learning algorithms to construct the KGML framework. Model performance was compared with that of standalone machine learning and process-based models across forest types and prediction horizons. Results: Incorporating the CO ₂ fertilization effect improved the predictive performance of the FCS model, increasing the coefficient of determination ( R ² ) from approximately 0.85 to above 0.90. The KGML model consistently outperformed the standalone machine learning and process-based models across the evaluated forest types. Its advantage was particularly evident for prediction horizons of 10 years or longer, over which the performance of conventional machine learning models declined substantially. Scenario simulations indicated that natural forests in Heilongjiang Province had the greatest projected carbon sequestration potential under SSP126. Depending on model configuration and forest type, projected carbon stocks in 2060 ranged from 1371.37 ± 422.04 to 1442.38 ± 316.77 Tg. Conclusion: Integrating ecological process knowledge with machine learning can improve the accuracy and long-term extrapolation capability of forest carbon stock predictions. The proposed KGML framework provides a potentially transferable approach for regional forest carbon assessment. Nevertheless, the projections remain subject to uncertainty arising from climate scenarios, process-model assumptions, forest inventory coverage, and the representation of disturbances and management effects.