Air pollution has been shown to have a direct effect on human health. In particular, PM_2.5 has been proven to be related to cardiovascular and respiratory problems. Therefore, it is important to have accurate models to predict high pollution events for this and other pollutants. We present different models that forecast PM_2.5 maximum concentrations using a Long Short-Term Memory (LSTM) based neural network and a Deep Feedforward Neural Network (DFFNN). Ten years of air pollution and meteorological measurements from the network of monitoring stations in the city of Santiago, Chile were used, focusing on the behaviour of three zones of the city. All missing values were rebuilt using a method based on discrete cosine transforms and photochemical predictors selected through unsupervised clustering. Deep learning techniques provide significant improvements compared to a traditional multi-layer neural networks, particularly the LSTM model configured with a 7-day memory window (synoptic scale of pollution patterns) can capture critical pollution events at sites with both primary and secondary air pollution problems. Furthermore, the LSTM model consistently outperform deterministic models currently used in Santiago, Chile.

空气污染已被证明对人类健康有直接影响。特别是，PM _2.5已被证明与心血管和呼吸系统问题有关。因此，重要的是要有准确的模型来预测这种污染物和其他污染物的高污染事件。我们展示了预测 PM _{2.5 的}不同模型使用基于长短期记忆 (LSTM) 的神经网络和深度前馈神经网络 (DFFNN) 的最大浓度。使用了智利圣地亚哥市监测站网络十年的空气污染和气象测量结果，重点关注该市三个区域的行为。使用基于离散余弦变换和通过无监督聚类选择的光化学预测因子的方法重建所有缺失值。与传统的多层神经网络相比，深度学习技术提供了显着的改进，特别是配置了 7 天记忆窗口（污染模式的天气尺度）的 LSTM 模型可以捕获具有初级和次级空气污染问题的地点的关键污染事件. 此外，