A Comparative Framework for Peak-Hour Traffic Accident Prediction with Explainable AI and Time- Series Models

Abstract—Rush-hour traffic accidents seriously impairurban mobility, cause enormous financial and human losses,and necessitate the use of accurate prediction techniques tomake preventative measures. The conventional statisticaltechniques are often hard to represent intricate spatialvariations, and non-uniform time relationships in the actualaccident data. To solve these issues, a time-series forecastingframework is developed on a publicly available AccidentPrediction dataset on Kaggle, to predict peak-hour trafficaccidents. Data preprocessing involves cleaning, normalizingthe data, categorical encoding, exploratory analysis, andfeature engineering based on location and peak-hour features.Some of the model prediction models applied include a longshort-term memory network to model long-term temporaltrends, a multilayer perceptron with lagged inputs, SARIMAXwith exogenous variables, and a Negative Binomial generalizedlinear model. The performance of the model is measured bymean error, mean absolute error, root mean square error,mean absolute percentage error and mean absolute scalederror. The results show that LSTM model makes the bestpredictions, the lowest MAE of about 1.35 and RMSE of about1.85. Explainability is added with LIME and SHAP, and aFlask-based interface allows users to add their ownpredictions, which can enhance the accuracy and usefulness ofpredictions in urban traffic safety