Spatiotemporal Graph Convolutional Recurrent Neural Network Model for Citywide Air Pollution Forecasting

TL;DR

Addresses citywide air pollution forecasting as a spatiotemporal problem and proposes a Spatiotemporal Graph Convolutional Recurrent Neural Network (ST-GCRNN) that fuses a Graph Convolutional Network with a GRU-based RNN to predict future graph signals $\tilde{X}^{(t+1:t+T')}$ from historical signals $X^{(t-T+1:t)}$ on a city-wide graph $G=(V,E,W)$. Uses a seq2seq encoder–decoder to forecast multiple steps and supports both spectral graph convolution (via Chebyshev polynomials on the normalized Laplacian) and diffusion convolution. Demonstrates superior short-term performance over ConvLSTM with a substantial reduction in parameters, and competitive long-term accuracy against a Hybrid GCN–LSTM baseline, while expanding the Seoul dataset from 2015–2017 to 2015–2019. Provides a compact, scalable approach for real-time urban air quality forecasting and contributes a large-scale data resource for spatiotemporal modeling.

Abstract

Citywide Air Pollution Forecasting tries to precisely predict the air quality multiple hours ahead for the entire city. This topic is challenged since air pollution varies in a spatiotemporal manner and depends on many complicated factors. Our previous research has solved the problem by considering the whole city as an image and leveraged a Convolutional Long Short-Term Memory (ConvLSTM) model to learn the spatiotemporal features. However, an image-based representation may not be ideal as air pollution and other impact factors have natural graph structures. In this research, we argue that a Graph Convolutional Network (GCN) can efficiently represent the spatial features of air quality readings in the whole city. Specially, we extend the ConvLSTM model to a Spatiotemporal Graph Convolutional Recurrent Neural Network (Spatiotemporal GCRNN) model by tightly integrating a GCN architecture into an RNN structure for efficient learning spatiotemporal characteristics of air quality values and their influential factors. Our extensive experiments prove the proposed model has a better performance compare to the state-of-the-art ConvLSTM model for air pollution predicting while the number of parameters is much smaller. Moreover, our approach is also superior to a hybrid GCN-based method in a real-world air pollution dataset.

Figures (11)

• Figure 1: A Graph structure constructed for the air pollution forecasting problem in Seoul city, South Korea. Each node denotes an air pollution monitoring station. The labels of edges are the real-world distances between stations (some are omitted).
• Figure 2: An air pollution forecasting task is constructed as a graph-based problem. $T$ is the number of historical graph signals, $T'$ is the number of future prediction graph signals. $X \in R^{N \times F}$ is an input graph signal, $\tilde{X} \in R^{N \times 1}$ is an output signal, $h(.)$ is the learned function given the graph $G$.
• Figure 3: The overall architecture of our Spatiotemporal Graph Convolutional Recurrent Neural Network model. Input Graph Signals are historical air pollution data represented in graph structures. Prediction Outputs are predicted air pollution values in some hours ahead. A GCRNN Layer includes a GRU cell with graph convolution operators. An Encoder and Decoder architecture is applied to predict multiple future time-steps.
• Figure 4: Modeling spatiotemporal impact factors by fusing all input data into a combined graph signal.
• Figure 5: Compare testing performance of ConvLSTM model (ConvLSTM) and Spatiotemporal GCRNN model (ST-GCRNN) for PM$_{10}$ and PM$_{2.5}$ air pollution forecasting from 1 to 12 hours in Seoul data (smaller is better).