Regional air quality monitoring, a critical component of sustainable development is realized through various air quality observation stations established across a region. Accurate forecasting of air quality data collected from these observation stations requires the modelling of spatial–temporal patterns in the data. Deep learning algorithms, known for their ability to capture layers of abstraction, can proficiently achieve spatial–temporal modeling. However, deterministic models that produces point forecast does not consider the underlying model uncertainty during prediction and are therefore less reliable for real-time applications. Probabilistic forecasting models that forecast prediction intervals rather than point estimates can overcome this through uncertainty quantification. The objective of the proposed study is three-fold: i) develop an efficient deterministic deep learning spatial–temporal neural network named DL-STNN for spatial–temporal air quality forecasting; ii) investigate different approaches to uncertainty quantification in deep learning models and integrate some of them, such as Monte-Carlo Dropout, Ensemble Averaging, Gaussian Process Regression, Quantile Regression, and Bayesian Inference, in tandem with DL-STNN to facilitate probabilistic forecasting; and iii) evaluate the developed deterministic and probabilistic models, using a real-world Delhi air quality dataset. The evaluation results show that, among the deterministic models, DL-STNN outperforms the baselines with 39.8% more accurate predictions and performs consistently across all seasons in Delhi. Furthermore, among the DL-STNN-based tandem models that performed probabilistic forecasting, Bayesian DL-STNN proved efficient. It does 13% more accurate point forecasting and has 20% higher suitability score than the other tandem models, indicating that Bayesian inference adapts DL-STNN more reliable for real-time applications.

区域空气质量监测是可持续发展的关键组成部分，通过在一个地区建立的各种空气质量观测站来实现。对从这些观测站收集的空气质量数据进行准确预测需要对数据中的时空模式进行建模。深度学习算法以其捕获抽象层的能力而闻名，可以熟练地实现时空建模。然而，产生点预测的确定性模型在预测过程中没有考虑潜在的模型不确定性，因此对于实时应用来说不太可靠。预测预测区间而不是点估计的概率预测模型可以通过不确定性量化来克服这一点。拟议研究的目标有三个：i) 开发一个名为 DL-STNN 的高效确定性深度学习时空神经网络，用于时空空气质量预测；ii) 研究深度学习模型中不确定性量化的不同方法，并将其中一些方法与 DL-STNN 结合使用，以促进概率预测；iii) 使用真实的德里空气质量数据集评估开发的确定性和概率模型。评估结果表明，在确定性模型中，DL-STNN 以 39.8% 的准确预测优于基线，并且在德里的所有季节都表现一致。此外，在执行概率预测的基于 DL-STNN 的串联模型中，贝叶斯 DL-STNN 被证明是有效的。与其他串联模型相比，它的点预测准确率提高了 13%，适用性得分提高了 20%，这表明贝叶斯推理使 DL-STNN 更可靠地适应实时应用程序。