Abstract: Objective: The Guangdong-Hong Kong-Macao Greater Bay Area is an important economic development demonstration zone in China and one of the mega-urban agglomerations most frequently affected by typhoons. In response to the intensifying risk of extreme typhoon disasters in this region under the background of global climate warming and urban expansion, as well as the shortcomings of traditional disaster assessment methods characterized by strong subjectivity and fragmented data sources, and the high computational cost and insufficient timeliness of high-fidelity physical models, this study aims to construct an efficient and accurate typhoon disaster loss risk assessment model to support regional quantitative risk management and disaster prevention resilience enhancement. Methods: Based on multi-source observational data, this study takes the comprehensive typhoon disaster assessment index as the research object, and selects four machine learning algorithms — Random Forest (RF), XGBoost, Artificial Neural Network (ANN), and LightGBM — to construct a typhoon disaster loss risk assessment model for the Greater Bay Area. Using historical typhoon disaster loss data from 2001 to 2023, the risk levels are classified and used as output variables to predict typhoon loss risks in the Greater Bay Area. Results: The research findings indicate that: 1) In terms of disaster loss regression, the Random Forest model achieves the best fitting performance, with a coefficient of determination R² of 0.93, and MSE and MAE values of 1.867 and 0.126, respectively, indicating small prediction bias and no extreme outlier errors. SHAP feature importance analysis reveals that gale duration and regional population size are the key factors affecting typhoon disaster losses. 2) In terms of disaster risk classification, the XGBoost model exhibits strong robustness in handling imbalanced data, achieving a macroaverage F1 score (reflecting class fairness) of 92.88% and an overall accuracy of 99.66%. Among the classes, the severe disaster category achieves an accuracy of 98% and a macro-average F1 score of 90%, demonstrating the model's strong robustness on class-imbalanced data. SHAP feature analysis shows that gale duration and total process rainfall contribute the most to the classification of disaster loss levels. Validation using three typical typhoons — Hato, Mangkhut, and Saola — yields an average prediction accuracy exceeding 96.8%, indicating high precision and good applicability of the model for disaster assessment in the Greater Bay Area. Conclusion: This study achieves a quantitative assessment of comprehensive typhoon risk levels in the Greater Bay Area, providing theoretical support and a scientific reference for refined regional typhoon risk management, precise allocation of emergency resources, and the enhancement of urban disaster prevention and resilience.