Autonomous vehicles face constraints from multiple factors, including insufficient technological maturity, complex urban road network structures, and behavioral uncertainties of traffic participants. Consequently, their driving safety has not yet been fully verified and guaranteed. To address the gaps in existing research, such as the lack of spatiotemporal dimension-specific prediction and insufficient utilization of higher-level autonomous vehicle accident data, a Geographically and Temporally Weighted Regression (GTWR)-HOLT accident risk prediction model was constructed. This model integrates accident severity and frequency while accounting for spatial heterogeneity and temporal non-stationarity. Based on collision accident data from autonomous vehicles in California, USA, for the period 2021-2023, and integrating multi-source data including road networks, Points of Interest (POI), and population, this study conducts empirical research by dividing the study area into spatiotemporal grids with a spatial step of 2 km × 2 km and a temporal step of one quarter. GTWR analysis of the first ten quarters of data shows a Mean Absolute Error (MAE) of 0.05, with predicted values closely matching the trend of the actual values, effectively capturing the variation patterns of risk. Commercial POI density, educational POI density, and population density consistently exhibit positive effects on accident risk. The risk prediction results for the eleventh quarter using the HOLT exponential smoothing model yield an MAE of 0.08. The results indicate that medium-high risk areas are likely to maintain their current risk levels in the near term. Predictions for low-to-medium risk areas and some high-risk areas perform well, and the predicted regression coefficients largely follow the trends of the original coefficients. The research provides data-driven support in formulating regulatory strategies for traffic management authorities and algorithm improvement measures in complex scenarios for vehicle manufacturers.