Periodic-CRN: A Convolutional Recurrent Model for Crowd Density Prediction with Recurring Periodic Patterns

Ali Zonoozi, Jung-Jae Kim, Xiao-Li Li, Gao CongIJCAI 2018 · DOI

Introduction

Predicting urban crowd density is a cornerstone of modern smart city initiatives, impacting everything from traffic management to the efficiency of ride-sharing platforms. While urban dynamics are complex and non-linear, they are also deeply rooted in human behavior, which is inherently periodic. Daily commutes, weekend leisure activities, and weekly work cycles create recurring patterns in crowd density that are highly predictable if modeled correctly.

Periodic patterns of taxi density over time in different urban regions.

This paper introduces Periodic-CRN (PCRN), a novel deep learning architecture designed to explicitly learn, store, and reuse memory representations of these recurring patterns. Unlike traditional models that treat each time step in isolation or rely on immediate historical context, PCRN utilizes a latent memory mechanism to capture long-term seasonality directly within its architecture.

The Challenge of Spatio-Temporal Forecasting

Spatio-temporal forecasting involves predicting future states based on historical data that has both spatial and temporal dimensions. Traditional linear models like ARIMA often fail to capture the complex, non-linear correlations found in city-scale traffic. Modern deep learning approaches, such as Convolutional LSTMs (ConvLSTM) or Convolutional GRUs (ConvGRU), have improved performance by capturing local spatial dependencies alongside temporal sequences.

However, standard recurrent neural networks (RNNs) possess a significant limitation: they struggle to explicitly "remember" long-term periodic patterns, such as the specific traffic conditions at 8:00 AM every Monday. Previous attempts to address this often involved feeding raw historical data from previous periods back into the model at runtime. This approach is computationally redundant and inefficient. PCRN seeks to overcome this by building a model that inherently remembers and dynamically utilizes these periodic patterns within its latent space.

Methodology

The PCRN architecture is built upon three primary innovations designed to optimize spatio-temporal learning and periodic memory retrieval.

The architecture of the Periodic-CRN (PCRN) model.

1. Pyramidal Convolutional GRU

To capture spatio-temporal correlations efficiently, the authors employ a pyramidal structure for the ConvGRU layers. Adapted from techniques in speech recognition, this structure involves concatenating consecutive outputs from lower layers before passing them to the upper layer. This effectively halves the sequence length at each step, which speeds up training and broadens the spatial receptive field without significantly increasing memory consumption.

2. Periodic Representations (Memory Dictionaries)

Rather than re-processing raw historical images, PCRN saves the hidden state tensors generated by the ConvGRU into specialized key-value dictionaries. These dictionaries act as a learned representation of specific periods, such as daily or weekly patterns. The model updates and retrieves these states based on a Temporal Ordering mechanism, which the authors found to be the most effective method for maintaining relevant historical context.

3. Weighting-Based Fusion

To integrate these various inputs, PCRN uses an attention-like weighting mechanism. This module calculates scalar weights for the daily periodic representation, the weekly representation, and the immediate context based on the current state. By dynamically fusing these historical tensors with current ConvGRU outputs and external meta-data (e.g., time of day, day of the week), the model produces a highly informed final prediction.

Experimental Evaluation

The model was rigorously evaluated on two major grid-based spatial datasets:

  • TaxiSG (Singapore): Comprising taxi trajectory logs from early 2015, aggregated into 5-minute intervals on a 32x32 grid of the Central Business District.
  • TaxiBJ (Beijing): A benchmark dataset tracking taxi in-flow and out-flow across a 32x32 grid of Beijing, aggregated into 30-minute intervals.

Key Results

PCRN consistently outperformed state-of-the-art baselines, including ST-ResNet and standard ConvGRU models, in both next-frame and multi-step forecasting tasks.

  • Next-Frame Accuracy: On the TaxiBJ dataset, PCRN achieved an RMSE of 15.85, significantly lower than ST-ResNet's 16.69. Similar gains were observed on the TaxiSG dataset.
  • Multi-Step Robustness: When predicting up to six time steps into the future, PCRN exhibited the lowest rate of error degradation. While linear models like ARIMA saw rapid error growth, PCRN maintained high accuracy, proving its effectiveness for longer-term urban planning.
  • Ablation Studies: The "Temporally Ordered" approach for updating memory dictionaries was shown to be superior to "Sequentially Ordered" or "Moving Average" methods, highlighting the importance of precise temporal alignment in periodic modeling.

Discussion and Conclusion

The success of PCRN suggests that storing learned hidden states (latent memory) is a far more effective strategy for modeling seasonality than relying on raw historical data. By using an attention-based weighting mechanism, the model can intelligently prioritize different historical periods based on the current urban context.

While the model currently relies on rigid 2D grid partitions—which may overlook the specific topology of road networks—it represents a significant step forward in periodic spatio-temporal forecasting. Future work may involve adapting these periodic memory mechanisms to graph-based architectures to better reflect the physical constraints of urban infrastructure.

Summary updated

BibTeX
@inproceedings{zonoozi2018periodic,
  author = {Ali Zonoozi and Jung-Jae Kim and Xiao-Li Li and Gao Cong},
  title = {{Periodic-CRN: A Convolutional Recurrent Model for Crowd Density Prediction with Recurring Periodic Patterns}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/519},
}
Figure