DeepSTCL: A Deep Spatio-temporal ConvLSTM for Travel Demand Prediction

Dongjie Wang, Yan Yang, Shangming Ning

Introduction

Accurate travel demand prediction is essential for the efficient operation of modern ride-hailing platforms like Didi and Uber. Precise forecasting allows for proactive vehicle dispatching, reducing passenger wait times and urban congestion. However, travel demand is inherently complex, exhibiting deep spatio-temporal dependencies. Traditional time-series models like ARIMA often fail to capture these non-linear relationships, while standard deep learning approaches typically focus on either temporal sequences (LSTMs) or spatial topologies (CNNs) in isolation. This paper introduces DeepSTCL, an end-to-end framework designed to bridge this gap by treating citywide demand as a continuous "video stream."

Methodology

The DeepSTCL (Deep Spatio-Temporal Convolutional LSTM) framework processes historical travel data through a multi-branch architecture to capture diverse human behavioral patterns.

The "Video" Metaphor

The city is discretized into an $n \times n$ geographical grid. The demand within each grid cell at a specific time interval is treated as a single pixel in an image frame. A sequence of these frames over time constitutes a video stream, allowing the model to leverage computer vision techniques for forecasting.

Three-Branch Architecture

To account for the cyclical nature of human travel, the model splits input data into three distinct temporal branches:

  1. Closeness: Captures recent trends (e.g., the last few hours).
  2. Period: Captures daily cycles (e.g., the same time yesterday).
  3. Trend: Captures weekly patterns (e.g., the same time last week).
The DeepSTCL framework architecture, illustrating the three-branch temporal processing (closeness, period, trend) and the integration of ConvLSTM layers for spatio-temporal fusion.

ConvLSTM Layers

Unlike standard LSTMs that require flattening spatial data—thereby losing geographic context—DeepSTCL utilizes Convolutional LSTMs (ConvLSTMs). By incorporating convolution operations directly within the recurrent cells, the network simultaneously learns temporal sequences and the spatial relationships between neighboring grid blocks. The outputs of these branches are eventually fused using a learned linear weighting to produce the final prediction map.

Datasets and Experimental Setup

The model was rigorously evaluated using the DIDI GAIA Order dataset, which contains ride-hailing records and GPS traces from Chengdu, China, collected in November 2016.

Discretization of the Chengdu urban area into a grid for spatio-temporal analysis.

For preprocessing, the city was divided into a $74 \times 74$ grid, with each cell representing approximately 1 km². Data was aggregated into 1-hour intervals. The researchers noted a critical trade-off: while higher-resolution grids (smaller cells) improve prediction accuracy by capturing local nuances, they significantly increase computational overhead due to the higher number of convolution operations.

Key Results

DeepSTCL was benchmarked against standalone LSTM (temporal-only) and CNN (spatial-only) models.

  • Quantitative Performance: DeepSTCL consistently outperformed all baselines across RMSE, MAE, and MAPE metrics. For instance, during the 9:00 AM rush hour, DeepSTCL achieved an RMSE of 3.16, significantly better than the CNN (3.65) and LSTM (4.23).
  • Temporal Stability: While baseline models showed performance degradation as the day progressed, DeepSTCL maintained stable, high-accuracy predictions throughout the entire 24-hour cycle.
  • Qualitative Fidelity: Heatmaps generated by the model closely matched ground-truth demand, accurately capturing the outward radiation of travel needs from the city center (e.g., Chun Xi Road) during peak hours.

Conclusion and Future Work

DeepSTCL demonstrates that framing spatio-temporal grid data as a video prediction task is a highly effective strategy for urban demand forecasting. By isolating closeness, period, and trend features, the model effectively reduces noise and focuses on relevant behavioral patterns. Future research will explore incremental learning to allow the model to adapt in real-time to sudden urban dynamics, such as accidents, extreme weather, or special events, without requiring full retraining.

Summary updated

BibTeX
@inproceedings{wang2018deepstcl,
  author = {Dongjie Wang and Yan Yang and Shangming Ning},
  title = {{DeepSTCL: A Deep Spatio-temporal ConvLSTM for Travel Demand Prediction}},
  booktitle = {IJCNN},
  year = {2018},
  url = {https://openreview.net/forum?id=eehE1vCd_T},
}
Figure