DSETA Improves Real-Time Travel Time Prediction in Dynamic Traffic

Predicting Estimated Time of Arrival (ETA) accurately is fundamental for intelligent transportation systems, especially in large, congested cities where traffic patterns are highly dynamic. Existing prediction models often struggle to adapt to sudden congestion or irregular events like holidays, leading to degraded performance. This research addresses these limitations by proposing DSETA, a Dual-Stage ETA prediction framework that employs continual learning. DSETA's innovation lies in its two-stage learning process: an "intra-day" stage that uses real-time data for immediate adaptation to short-term events, and an "inter-day" stage that leverages aggregated historical data to capture long-term distribution shifts such as seasonal trends. To prevent "catastrophic forgetting" of regular patterns, a Historical Traffic Knowledge Consolidation module is integrated. Extensive offline and online experiments on DiDi's platform across major Chinese cities demonstrated significant Mean Absolute Error (MAE) reductions, confirming DSETA's effectiveness and robustness. The framework is already deployed in production, handling hundreds of millions of daily requests.