AI Predicts Risky Driving Hotspots Using Connected Vehicle Data

Historically, road safety initiatives have been reactive, primarily relying on analyzing crash records after incidents have already occurred. A critical challenge in this field is the proactive identification of high-risk locations and dangerous driving behaviors *before* accidents happen. This research tackles this gap by leveraging connected vehicle telemetry data collected from Greater Sydney, Australia. The study quantifies risky driving using g-force thresholds for events like hard braking, harsh cornering, and harsh acceleration. Spatio-temporal heatmaps are then constructed to pinpoint high-risk zones within Local Government Areas (LGAs). To forecast these near-miss risky driving events, eight different predictive models were benchmarked, spanning ensemble learning (Random Forests, XGBoost), deep learning (LSTM, N-BEATS), and classical time-series methods (ARIMA, Exponential Smoothing, Prophet). The results showed that ARIMA, a classical time-series model, achieved the lowest mean absolute error, performing comparably to LSTM and outperforming ensemble methods. This suggests that simpler, parsimonious time-series models can be highly effective, especially when training data volume is limited. The study successfully identified inner and western Sydney LGAs (CBD, Parramatta, Bankstown) as persistent high-risk zones, underscoring the significant potential of IoT-based connected vehicle data to inform and support proactive road safety policies and interventions.