To establish a "mechanism-constrained and data-driven" rockburst monitoring and early warning system, artificial intelligence algorithms are employed to construct nonlinear relational models that deeply explore inherent data patterns, thereby achieving an integrated "location-time-intensity" tripartite early warning framework. A multi-parameter rockburst risk prediction index system was established, incorporating real-time monitoring indicators, mining technical indicators, coal-rock mass strength indicators, and geological condition indicators. By introducing temporal sequence data and spatial location parameters, a spatiotemporally coupled early warning module was developed. Key methodologies include: scalar normalization of indicator parameters using normalization equations; K-means clustering to impute missing target values; correlation analysis and principal component analysis (PCA) to investigate inter-factor relationships and reveal mappings between data and geological/mining conditions, thereby unifying the "mechanism" essence with "data" phenomena. Three multi-parameter comprehensive early warning models—L2-regularized multiple linear regression, BPNN (Backpropagation Neural Network), and RNN-GRU (Recurrent Neural Network with Gated Recurrent Units)—were constructed and comparatively evaluated. Validation results demonstrate that the L2-regularized linear regression model effectively outputs fitting relationships between factors and rockburst risks while providing weight values of indicators for mechanistic studies. In contrast, the RNN-GRU model exhibits superior computational efficiency, faster convergence, and enhanced capability to capture dynamic features in time-series data through its GRU architecture. By integrating One-Hot encoding of monitoring point coordinates and improved K-means clustering for spatial risk localization, the model achieves spatiotemporal distribution prediction of rockburst hazards. Field validation at a rockburst-prone mine in Hujierte Mining District demonstrated, predicted magnitudes closely benchmarked against actual events, achieving <2.5-hour temporal MAE and <50-meter spatial RMSE, with spatiotemporal prediction reliability reaching 85% confidence level.