Abstract
<jats:p>The occurrence of forest fires in China was accurately predicted to optimize resource allocation and preventive mitigation. Drawing on climate, forest resource, and socio–economic factors together with forest fire incident records from 2003 to 2023, we construct a deep learning model that integrates a Convolutional Neural Network - Gated Recurrent Unit (CNN-GRU) with a Multi-Head Attention (MHA) mechanism and a Kepler Optimization Algorithm (KOA). The CNN-GRU captures spatiotemporal features, MHA enhances the recognition of intrinsic data relationships, and KOA automatically tunes network parameters. Our KOA-CNN-GRU-MHA model surpasses traditional machine learning baselines and the native CNN-GRU model, reducing the mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) by 26.75%, 64.35%, and 16.47%, respectively, for the total annual Number of Forest Fires (NFF), and by 39.07%, 76.03%, and 32.82% for the total annual Number of Small Forest Fires (NSFF). Ranking the influence of predictor variables further guides fire management strategies and supports more effective operational planning.</jats:p>