Air pollution is one of major environmental issues in the twenty-first century due to global industrialization and urbanization. Its mitigation necessitates accurate air quality forecasts. However, current state-of-the-art air quality forecasts are limited from highly uncertain chemistry-transport models (CTMs), shallow statistical methods, and heterogeneous and incomplete observing networks. Recently, deep learning has emerged as a general-purpose technology to extract complex knowledge using massive amount of data and very large networks of neurons and thus has the potential to break the limits of air quality forecasts. Here, we provide a brief review of recent attempts on using deep learning techniques in air quality forecasts. We first introduce architectures of deep networks (e.g., convolutional neural networks, recurrent neural networks, long short-term memory neural networks, and spatiotemporal deep network) and their relevance to explore the nonlinear spatiotemporal features across multiple scales of air pollution. We then examine the potential of deep learning techniques for air quality forecasts in diverse aspects, namely, data gap filling, prediction algorithms, improvements of CTMs, estimations with satellite data, and source estimations for atmospheric dispersion forecasts. Finally, we point out some prospects and challenges for future attempts on improving air quality forecasts using deep learning techniques.

由于全球工业化和城市化，空气污染是21世纪的主要环境问题之一。为了缓解这种状况，必须进行准确的空气质量预测。但是，当前的最新空气质量预报受到高度不确定的化学运输模型（CTM），浅层统计方法以及异构和不完整的观测网络的限制。近年来，深度学习已成为一种通用技术，可以利用大量数据和非常大的神经元网络来提取复杂的知识，因此有可能突破空气质量预报的极限。在这里，我们简要回顾了最近在空气质量预测中使用深度学习技术的尝试。我们首先介绍深度网络的架构（例如卷积神经网络，递归神经网络，长短期记忆神经网络和时空深层网络）及其相关性，以探讨跨多个尺度的空气污染的非线性时空特征。然后，我们在各个方面检查深度学习技术在空气质量预测中的潜力，即数据缺口填充，预测算法，CTM的改进，卫星数据的估计以及大气扩散预测的源估计。最后，我们指出了使用深度学习技术改进空气质量预报的未来前景和挑战。然后，我们在各个方面检查深度学习技术在空气质量预测中的潜力，即数据缺口填充，预测算法，CTM的改进，卫星数据的估计以及大气扩散预测的源估计。最后，我们指出了使用深度学习技术改进空气质量预报的未来前景和挑战。然后，我们在各个方面检查深度学习技术在空气质量预测中的潜力，即数据缺口填充，预测算法，CTM的改进，卫星数据的估计以及大气扩散预测的源估计。最后，我们指出了使用深度学习技术改进空气质量预报的未来前景和挑战。