Abstract: This research paper tackles the challenges of accuracy and interpretability in forest fire prediction studies. The study focuses on the Yunnan province as the geographical area of interest and proposes an innovative approach that combines the LightGBM machine learning model with the SHAP (Shapley Additive explanations) interpretable model for forest fire prediction. The findings indicate that compared to other existing machine learning models, LightGBM demonstrates superior accuracy, achieving a rate of 90.5% in accurately identifying most fire incidents within high?risk zones. By incorporating the SHAP model, the interpretability of the LightGBM model is enhanced. The global interpretability method aids users in comprehending the response patterns of various features within the model, while the local interpretability method elucidates the prediction process for specific instances of forest fires. The proposed methodology, utilizing LightGBM and SHAP, not only advances forest fire prediction in Yunnan province but also offers valuable insights for risk prediction modeling in other domains of disaster forecasting.

Key words: forest fires, predictive models, machine learning, interpretability