一种空气质量预测模型的研究

发布时间：2018-03-26 20:25

本文选题：空气质量预测　切入点：遗传算法　出处：《南京邮电大学》2017年硕士论文

【摘要】：空气污染有碍人类健康,近年来人们对于空气质量预测的渴望有增无减。随着空气质量检测系统的大量建立,相关数据的变得逐渐丰富起来,这使得预测空气质量成为可能。本文根据南京市从2015年10月到2016年8月,10个月以来的空气污染浓度数据及气象数据作为研究的数据集,综合考虑了涉及空气质量计算的六种有害物质的当前浓度以及它们所对应的前一时刻的浓度、温度、相对湿度、天气形势、大气压、降水、季节、是否交通高峰时段、风力与所处的风向这21个特征维度。并通过主成分分析法对数据在最大程度保留的基础上进行了降维,有效的缩减了输入数据的规模。以此为基础,提出且证明了一种空气质量模型的可行性。该模型的主体仍然基于BP神经网络,但是对BP神经网络易于陷入局部极小值、收敛速度慢、预测精确程度不高的缺点进行了改良。模型的特点在于使用遗传算法对参数进行了寻优,使用了基于贝叶斯正则化改进的Levenberg-Marquardt反向传播算法(L-M反向传播算法)以及启用了双隐层设计思路。论文论证了所提出模型是切实可行的,并且在与传统BP神经网络空气质量模型实验比较中证明了新的模型在收敛速度、训练精度、预测能力这些方面的提升十分显著。
[Abstract]:Air pollution is harmful to human health, and the desire for air quality prediction has increased in recent years. With the establishment of a large number of air quality monitoring systems, the relevant data have become more and more abundant. This makes it possible to predict air quality. Based on data on air pollution concentrations and meteorological data from October 2015 to August 2016 in Nanjing, Taking into account the current concentrations of the six harmful substances involved in the calculation of air quality and their corresponding concentrations at the previous moment, temperature, relative humidity, weather situation, atmospheric pressure, precipitation, seasons, whether the traffic is in peak time, Wind force and wind direction are 21 characteristic dimensions. The dimension of data is reduced on the basis of maximum retention by principal component analysis, which effectively reduces the size of input data. The feasibility of an air quality model is presented and proved. The main body of the model is still based on BP neural network, but the BP neural network is prone to fall into a local minimum, and the convergence rate is slow. The model is characterized by the use of genetic algorithms to optimize the parameters. The improved Levenberg-Marquardt backpropagation algorithm based on Bayesian regularization is used, and the design idea of double hidden layer is enabled. The paper proves that the proposed model is feasible. Compared with the traditional BP neural network air quality model, it is proved that the new model can improve the convergence rate, training precision and prediction ability.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：X51;TP18

【参考文献】