Boosting ETA Accuracy: How TCN Improves Historical Speed Prediction for Navigation

Introduction

Accurate estimated time of arrival (ETA) is a core metric for navigation services. Users typically compare three routes displayed by the front‑end and rely heavily on the predicted travel time, which is derived from real‑time traffic conditions and historical speed information.

Problem with Historical Average Method

The conventional approach computes historical speed by averaging travel times of the same feature day and time slot across past weeks. This assumes "history equals future" and works for short‑term fluctuations but fails for longer horizons where annual seasonality and trends dominate. The method is also sensitive to outliers, cannot capture temporal trends, and ignores year‑over‑year patterns, leading to significant errors in ETA for long routes.

Proposed Machine‑Learning Solution

A Temporal Convolutional Network (TCN) is employed to model historical speed more accurately. The model ingests both dynamic and static features to predict the average travel time for the next week.

TCN Overview

TCN is a convolution‑based sequence modeling architecture with causal convolutions, dilated kernels, and residual blocks, enabling long‑range dependency capture without information leakage from future to past.

Network Architecture

The overall framework consists of a dynamic feature extraction module (TCN) and a static feature module. Dynamic features are learned from a dual‑channel input comprising this year’s and last year’s average travel time sequences. Static features are concatenated with the dynamic output before the final regression layer.

Dynamic Feature Extraction

The TCN learns patterns from sequences of average travel times for the current and previous year, capturing trends, seasonality, and recent anomalies. Experiments comparing RNN, LSTM, and TCN showed that TCN achieved the best performance, improving accuracy by 1.39% over RNN and 0.83% over LSTM while training faster.

Static Features

Road attributes: length, width, number of lanes, lane width, maximum speed limit.

Time attributes: travel times of the same time slot in the past three days, average travel time of the past seven days, and the two average travel times from the same period last year (both before and after the target slot).

Road attributes reflect differing capacities, while time attributes capture recent conditions and yearly periodic effects.

Model Effects

Automatic Outlier Filtering

The TCN effectively ignores abnormal spikes (e.g., accidents) that would otherwise inflate the historical average, resulting in more realistic ETA predictions.

Trend Extraction

By learning the upward trend in recent travel times, the model predicts a continued increase for the following week, whereas the historical average underestimates due to earlier low values.

Incorporating Annual Seasonality

When a sudden rise occurs in week 12, the TCN leverages last year’s corresponding week data to anticipate a similar rise in week 13, addressing cases where the historical average would miss the seasonal pattern.

Evaluation Results

On a weekly case set, the baseline historical average method exhibited a bad‑case rate of 11.0‰. The TCN‑based approach reduced this rate to 10.1‰, demonstrating a notable improvement by handling annual cycles and outliers.

Conclusion

The industrial deployment of a TCN model, combined with engineered dynamic and static features, successfully mitigates the deficiencies of the historical average method, improves ETA accuracy, and provides a viable pathway for tackling other time‑series problems in navigation.