Long short-term memory neural network for air pollutant concentration predictions: Method development and evaluation
Long short-term memory neural network for air pollutant concentration predictions: Method development and evaluation
复制标题
DOI：
10.1016/j.envpol.2017.08.114
复制
发表时间：
2017-12-01
影响因子：
8.9
通讯作者：
Chi, Tianhe
中科院分区：
文献类型：
--
作者：
Li, Xiang;Peng, Ling;Chi, Tianhe
空气污染物的集中度预测是一种有效的方法，可以通过对有害空气污染物的预警预警。但是，现有的空气污染物浓度预测方法无法有效地模拟长期依赖性和大多数忽视的空间相关性。在本文中，提出了一种新型的长期记忆神经网络扩展（LSTME）模型，该模型固有地考虑了空气污染物浓度预测的时空相关性。长期记忆（LSTM）层用于自动从历史空气污染物数据中提取固有的有用特征，并将辅助数据（包括气象数据和时间邮票数据）合并到提出的模型中以增强性能。每小时PM2.5（空气动力学直径小于或等于2.5 mu m的颗粒物物质）从2014年1月/2014年1月至5月/2016年5月28日在北京市12个空气质量监测站收集的浓度数据用于验证有效性提出的LSTME模型。使用时空深度学习（STDL）模型，时间延迟神经网络（TDNN）模型，自动回归移动平均值（ARMA）模型，支持向量回归（SVR）模型以及传统的LSTM NN模型以及A结果的比较表明，LSTME模型优于其他基于统计的模型。此外，使用辅助数据改善了模型性能。对于一个小时的预测任务，提出的模型表现良好，平均绝对百分比误差（MAPE）为11.93％。此外，即使在13-24 h预测任务（MAPE = 31.47％）中，我们在不同时间跨度上进行了多尺度预测，并达到了令人满意的性能。 （c）2017 Elsevier Ltd.保留所有权利。
Air pollutant concentration forecasting is an effective method of protecting public health by providing an early warning against harmful air pollutants. However, existing methods of air pollutant concentration prediction fail to effectively model long-term dependencies, and most neglect spatial correlations. In this paper, a novel long short-term memory neural network extended (LSTME) model that inherently considers spatiotemporal correlations is proposed for air pollutant concentration prediction. Long short-term memory (LSTM) layers were used to automatically extract inherent useful features from historical air pollutant data, and auxiliary data, including meteorological data and time stamp data, were merged into the proposed model to enhance the performance. Hourly PM2.5 (particulate matter with an aerodynamic diameter less than or equal to 2.5 mu m) concentration data collected at 12 air quality monitoring stations in Beijing City from Jan/01/2014 to May/28/2016 were used to validate the effectiveness of the proposed LSTME model. Experiments were performed using the spatiotemporal deep learning (STDL) model, the time delay neural network (TDNN) model, the autoregressive moving average (ARMA) model, the support vector regression (SVR) model, and the traditional LSTM NN model, and a comparison of the results demonstrated that the LSTME model is superior to the other statistics-based models. Additionally, the use of auxiliary data improved model performance. For the one-hour prediction tasks, the proposed model performed well and exhibited a mean absolute percentage error (MAPE) of 11.93%. In addition, we conducted multiscale predictions over different time spans and achieved satisfactory performance, even for 13-24 h prediction tasks (MAPE = 31.47%). (C) 2017 Elsevier Ltd. All rights reserved.