How Gaode’s ETA System Predicts Arrival Times with Hybrid Spatio‑Temporal GCN

1. What Is Driving ETA

ETA (Estimated Time of Arrival) indicates the expected travel duration from the current moment to a destination along a given route. In Gaode Maps, ETA is displayed on the front‑end navigation page and influences user trip planning, ride‑hailing pricing, and various business scenarios.

Key technical concepts include:

link : the basic road segment between two intersections.

trace : a sequence of links representing a vehicle’s trajectory.

SP (Speed Profile) : historical average speed for a link during a specific time window, aggregated in 10‑minute intervals.

AutoLR (Autonavi Location Reference) : real‑time travel time for a link, updated every minute.

Link travel‑time ground truth : average travel time of all traces that entered a link within a 5‑minute window after noise removal.

2. Business Components

The ETA service consists of several layers:

Data layer : Links are categorized based on historical ground‑truth travel times. Some links, which rarely experience congestion, can be predicted using static SP alone, while others require more sophisticated models.

Prediction modules :

Static SP – uses historical averages without real‑time data.

Linear model – combines real‑time AutoLR and historical SP to fit link‑level ground truth.

Precise prediction – a deep‑learning model that leverages real‑time speed, flow, and other signals for complex, sudden congestion.

Prediction functions : Provide link‑level forecasts, trace‑level (whole‑route) ETA, and punctuality probability.

Application layer : Both self‑driving and ride‑hailing services consume ETA, with different sample distributions and focus areas.

Evaluation & monitoring : Includes link‑level and trace‑level metric assessments, real‑time issue detection, and automated or manual interventions.

3. ETA Prediction Workflow

The online ETA service offers an API for other services. Precise prediction results are cached every minute by a separate algorithm service. When a link is queried, the system prefers precise predictions; if unavailable, it checks whether real‑time AutoLR data exists and whether the prediction horizon is within one hour. If both conditions hold, the linear model is used; otherwise, static SP is applied. SP data, being historical averages, can support forecasts up to a week ahead.

For whole‑route ETA, the system iteratively aggregates link‑level predictions along the planned trace, advancing the time window after each link. If cumulative time exceeds one hour, static SP is used for remaining links.

4. Challenges and Solutions

High‑traffic, highly variable scenarios (e.g., weekday rush hours, holiday travel spikes, special dates) make accurate ETA prediction difficult because historical patterns may not apply.

To enable the model to anticipate future congestion, a “future traffic flow” feature is introduced. By aggregating the planned routes of users who have already departed, the system can estimate the traffic volume that will appear on each link later in the day, even before congestion materializes.

The proposed deep‑learning framework, Hybrid Spatio‑Temporal Graph Convolutional Network (H‑STGCN) , incorporates the future traffic flow feature. It uses a domain transformer to fuse heterogeneous traffic modalities and a composite adjacency matrix to better capture spatial relationships between links. Experiments on real‑world datasets show that H‑STGCN outperforms state‑of‑the‑art spatio‑temporal models, especially for sudden congestion, and the work was accepted at KDD 2020.

5. Evaluation

5.1 Link‑Level Metrics

Two primary metrics are used:

Bad link rate : proportion of link batches classified as “bad”. Lower values indicate better performance.

MSE (Mean Squared Error) : computed over all links (n) and evaluation batches (k). Smaller MSE reflects higher accuracy.

5.2 Trace‑Level Metrics

Trace‑level evaluation compares the user’s actual travel time with the predicted ETA, reporting MSE, MAE, MAPE, as well as good‑rate and bad‑rate indicators.

6. Conclusion

The article provides an overview of Gaode’s driving ETA system, its data and prediction layers, and introduces the H‑STGCN model that leverages future traffic flow for more accurate spatio‑temporal forecasting. The associated research paper was presented at KDD 2020.