Drift3Flow: Freeway-Incident Prediction Using Real-Time Learning

Traffic congestion is a major problem on today´s urban mobility. This paper introduces a novel model for Automatic Incident Prediction (AID) on freeways: Drift3Flow. This stepwise methodology produces flow/occupancy rate predictions using an online weighted ensemble schema of two well-known time series analysis techniques: Autoregressive Integrated Moving Average (ARIMA) and Holt-Winters Exponential Smoothing (ETS). Then, it continuously monitors the probability distribution function (p.d.f.) of the prediction residuals to trigger alarms of an imminent prediction divergence, i.e. concept drift. Such alarm activates an update neuron which extends our model´s reactivity by embedding a fully incremental learning schema inspired on the Delta Rule (DR) (derived from the BackPropagation (BP) algorithm). Our experimental test-bed used three weeks of data acquired from a real-world sensor network in Asia. The results validated its contributions by exhibiting a superior performance: 25% greater than the one obtained using ARIMA and ETS-based AID methods.