مدل‌سازی مکانی–زمانی و پیش‌بینی نرخ وقوع تصادفات در مناطق شهری با استفاده از مدل رگرسیون وزنی مکانی- زمانی (GTWR)

Introduction: Urban traffic safety is a critical challenge in metropolitan areas, where accidents are influenced by factors like travel demand, vehicle ownership, and road networks. Traditional regression models often assume spatial stationarity, limiting their ability to capture spatial and temporal heterogeneity in accident patterns. To address this, this study applies a Geographically and Temporally Weighted Regression (GTWR) model to analyze and predict accident rates across Traffic Analysis Zones (TAZs) in Mashhad, Iran. The primary aim is to evaluate the local spatio-temporal effects of key variables and compare GTWR’s performance against a global regression model.

Materials and Methods: For this study, four years of traffic accident data (2021–2024) were collected and standardized for 253 TAZs in Mashhad. This was combined with independent variables related to travel demand (vehicle kilometers traveled and the number of generated trips), vehicle ownership, and infrastructure (number of unsignalized intersections). A global regression model was estimated first, followed by the calibration of the GTWR model. Model development used data from the first three years (2021–2023), with 70% randomly selected for training and calibration. The coefficient of determination (R²) and Residual Sum of Squares (RSS) were used to evaluate calibration. The remaining 30% of the data from these years were used for testing and validation, with model accuracy assessed using the correlation coefficient (R²). To evaluate predictive performance, data from 2024 was employed, and results were examined using the Pearson correlation coefficient (R²) and Mean Squared Error (MSE). Additionally, the bivariate Moran’s I index was applied to examine spatio-temporal autocorrelation in the original data and the model residuals.

Results and Discussion: Preliminary data analysis indicated that the traffic accident rate exhibited significant positive spatio-temporal autocorrelation, highlighting the necessity of using local models like GTWR. The findings demonstrate that the GTWR model significantly outperformed the global regression model. Specifically, the R² increased from 0.38 in the global model to 0.87 in the local model, while the RSS values showed a substantial reduction (62935.6 vs. 654567.76). Validation using the testing dataset confirmed this, with an R² of 0.83 for GTWR compared to 0.71 for the global model. Furthermore, predictions for 2024 revealed that GTWR achieved superior predictive performance, evidenced by a lower MSE (1.28 vs. 8.4) and a higher correlation coefficient (0.84 vs. 0.64) relative to the global model. The spatially varying local coefficients indicated pronounced spatio-temporal non-stationarity. For instance, the effect of vehicle kilometers traveled (VKT) was stronger in central and northern parts of the city, while weaker effects were observed in the southern zones. The automobile ownership (ACO) variable exhibited a positive and stable association with accident rates, particularly in central and high-density areas. Conversely, the number of unsignalized intersections showed stronger effects in the western and northwestern parts. The number of generated trips had a positive impact across all zones but exhibited greater spatio-temporal stability compared to the other variables. The bivariate Moran’s I results indicated that the residuals of the global model still suffered from spatio-temporal autocorrelation, whereas this issue was effectively mitigated in the GTWR model.

Conclusion: The results demonstrate that in urban traffic accident analysis, using local spatio-temporal models like GTWR can provide a more accurate representation of influencing factors by capturing spatial and temporal heterogeneity. Compared to global models, GTWR not only achieved a better fit but also revealed localized patterns in the effects of explanatory variables. These findings confirm the importance of travel demand and roadway infrastructure in urban traffic accidents and can serve as a foundation for targeted safety policies. Based on the results, focusing on urban travel demand management, controlling private car ownership and usage in high-density areas, and improving the safety of unsignalized intersections, particularly in the western and northwestern parts of the city, can help reduce accident rates. Ultimately, this study shows that, in addition to higher predictive accuracy, the GTWR model is an effective tool for prioritizing high-risk areas and optimizing the allocation of safety resources in urban planning.