Machine Learning Approaches for Predicting Peak-Hour Urban Traffic Accidents

Abstract

Traffic accidents during rush hours are a big problem for urban transportation. They kill people and cost a lot of money, so we need accurate predicting tools to help us make smart decisions before they happen. Real-life accident data often has nonlinear temporal dependencies and complex spatial variations that are hard for traditional statistical methods to explain. To solve these problems, a system for time-series forecasting is created to predict traffic accidents during rush hour using the Accident Prediction dataset from Kaggle, which is free for everyone to use. The data go through preprocessing steps like cleaning, normalization, exploratory analysis, categorical encoding, and feature engineering based on location and peak-hour characteristics. There are several prediction models used, such as a Negative Binomial generalized linear model, SARIMAX with external variables, a multilayer perceptron with lagged inputs, and a long short-term memory network for capturing trends over long periods of time. Mean error, mean absolute error, root mean square error, mean absolute percentage error, and mean absolute scaled error are all ways to measure how well a model works. The results show that the LSTM model does a better job of predicting the future, with an MAE of about 1.35 and an RMSE of about 1.85. LIME and SHAP are used to add explainability, and a Flask-based interface lets users make predictions, which improves the accuracy of predictions and makes them more useful for improving road safety in cities