【作者】 张蕊；

【导师】 赵小松；

【作者基本信息】 天津大学 ， 管理科学与工程， 2018， 硕士

【摘要】 空气质量问题近些年来已经引起全世界范围的关注。空气污染对于人类健康有着非常严重的影响。许多城市都存在着非常严重的空气污染问题。然而,实时检测空气质量的成本较高,而且实际操作难度大。因此,有效的空气质量预测方法对于保护人类健康、提高人类幸福感有很重要的意义。本文将应用深度学习的方法来预测美国三个著名工业城市空气质量的分类。本文将建立的递归神经网络模型(Recurrent Neural Network)作为主要预测工具。递归神经网络模型可以有次序的处理和记忆序列数据,例如一段时间内的每日空气质量数据。本文展示了递归神经网络模型,以及两种传统的机器学习方法,包括支持向量机模型(Support Vector Machine)和随机森林模型(Random Forest)一共三种网络模型的预测空气质量分类的实验结果。结果表明,本文提出的递归神经网络模型与其余两种方法相比,在不同的特征组合和序列长度实验中,具有更好的预测效果。相比于使用污染物的浓度,使用六种污染物的空气质量分指数值作为输入变量可以更显著的提高预测准确率。其中,六种污染物的浓度和全天的空气质量总指数的特征组合在预测空气质量分类问题上最有效。此外,递归神经网络模型在预测空气质量分类时,相比具有普通结构的一般模型,充分考虑了数据随着时间序列变化的动态属性,更能发挥其递归结构的效果。更多还原

【Abstract】 Having attracted attention worldwide,air pollutions are considered to have detrimental effects on human health.Many cities have suffered severe air pollution.However,monitoring the real-time air quality expenses much and the operation process is very difficult.Forecasting performance of air quality,thus,becomes an important issue for the welfare of people.In this research,we attempt to use a deep learning method to predict Air Quality Classification(AQC)on three famous industrial cities in United States.The Recurrent Neural Network(RNN)of deep learning is used to build a major prediction model.RNN can process and memorize the sequential data such as data concerning daily air quality in a given period of time.The experimental results show the performances on three models including Support Vector Machine,Random Forest and RNN.Our proposed RNN model has best results compared with two machine learning approaches in different experiments with different feature sets and data lengths.Using individual air quality indexes of six pollutants as input variables makes accuracy higher compared with those of concentrations.Concentrations of six pollutants combined with air quality index are considered as the most efficient feature sets.In addition,the sequential data on air quality problem used by RNN with memory model outperforms without memory operation for the reason that the dynamic attributes of data are fully taken into consideration.更多还原