Integrating Statistical, Machine Learning, and Deep Learning Approaches for Urban Traffic Accident Forecasting

Abstract

During rush hours, urban mobility is significantly impacted by traffic accidents, incurring significant economic and human costs, and requiring accurate forecasting tools to make proactive decisions. The complex spatial and temporal structures observed in real-world accident data are often complex and challenging to model using traditional statistical techniques. To address these problems, an Accident Prediction dataset from Kaggle is used as a publicly accessible dataset and a time-series forecasting framework is created for predicting the peaks of traffic accidents. The proposed preprocessing steps are dependent on location and peak hour characteristics, and involve cleaning, normalization, categorical encoding, exploratory analysis, and feature engineering. A number of prediction models are used, such as a long short-term memory network to capture longrange temporal patterns, a multilayer perceptron with lagged inputs, SARIMAX with exogenous variables, and a Negative Binomial generalized linear model. The model's performance is evaluated using mean error, mean absolute error, root mean square error, mean absolute percentage error and mean absolute scaled error. Based on the results, the LSTM model has the lowest MAE, approximately 1.35, and RMSE, approximately 1.85, making it the most accurate in prediction. LIME and SHAP are added to enhance the explainability, while the addition of a user-driven forecasting interface using Flask contributes to increased accuracy and usefulness of the prediction in the field of urban traffic safety.