城市环境噪声污染对人体健康的危害日益严重。城市交通以及商业、文化集会等城市活动所产生的高频噪声具有典型瞬时性、分散性、随机性等特征，已有区域性监测网络和稀疏周期监测频次难以有效支撑城市噪声污染的精细化管控。本文围绕城市环境噪声动态模拟这一热点领域，运用大数据思维和机器学习方法，基于多源大数据构建了高时空精度的城市环境噪声动态模拟模型，并在此基础上建立了城市环境噪声地图仿真技术平台，为城市环境噪声的精细化监测和动态化管控提供研究支撑。本文以北京海淀区为案例研究区域，在获取固定噪声自动监测站数据基础上，通过行走观测、车载观测两种方式，分别获取了高时空分辨率的移动噪声大数据，结合城市兴趣点、路网、土地利用数据、实时交通等声源大数据，建立城市环境噪声数据库，并对多源大数据进行了清洗、标准化和数据融合。论文运用基于移动时间窗口的神经网络、卷积神经网络、循环神经网络等三种机器学习方法开展城市环境噪声建模研究，对比了不同算法、不同数据源模型的模拟精度，建立融合了固定噪声源数据模型与移动噪声源数据模型的城市环境噪声地图仿真系统，实现了高时空精度城市环境噪声的动态模拟与分析识别。案例研究结果表明，对固定噪声自动监测数据的建模分析表明，相比移动时间窗口神经网络、卷积神经网络模型，循环神经网络模型模拟结果最优，训练数据集R值达0.92，验证数据集R值达0.92，声强级平均误差为3.4分贝；对移动噪声监测数据建模分析表明，训练、验证数据集R值分别为0.96、0.91，平均误差为4.3分贝。运用以上两个噪声大数据模型分别模拟区域环境噪声和道路交通噪声，并采用时空数据填补技术，将模拟结果放大到整个城市地区，建立了空间精度为200米*200米、时间精度为分钟级的城市环境噪声仿真地图。模拟结果表明，2018年6月1日-6月3日期间，海淀区共有4.24平方千米区域存在不同时刻噪声声强级超标，空间上主要集中在五道口、中关村周边，时间上集中在9:00、18:00、19:00等通勤时段，分别对应噪声超标区域面积为0.64平方千米、0.54平方千米和0.58平方千米。

Urban environmental noise pollution is increasingly harmful to human health. The high-frequency noise generated by urban activities such as urban traffic, commercial and cultural gatherings has the characteristics of typical instantaneity, dispersion and randomness. Existing regional monitoring networks and sparse periodic monitoring frequencies are difficult to effectively support the fine control of urban noise pollution. This paper focuses on the hot field of urban environmental noise dynamic simulation, using big data thinking and machine learning methods. Based on multi-source big data, a high-precision dynamic simulation model of urban environmental noise is constructed, and on this basis, the map simulation technology platform of urban environmental noise is established, which provides research support for the fine monitoring and dynamic control of urban environmental noise.This paper takes Haidian District of Beijing as a case study area. On the basis of obtaining the data of fixed automatic noise monitoring station, the mobile noise big data with high spatio-temporal resolution is obtained through walking observation and vehicle observation. Combined with the big data of urban interest points, road network, land use data, real-time traffic and other sound sources, the urban environmental noise database is established. Subsequently, the cleaning, standardization and data fusion of multi-source big data were completed. In this paper, three kinds of machine learning methods based on moving time window neural network, convolution neural network and cyclic neural network are used to study the urban environmental noise modeling. By comparing the simulation accuracy of different algorithms and different data source models, the urban environment noise map simulation system that integrates the fixed noise source data model and the mobile noise source data model is established to realize the dynamic simulation and analysis of urban environmental noise with high spatial and temporal accuracy.The results of the case study are as follows. The modeling analysis of fixed noise automatic monitoring data shows that, compared with moving time window neural network and convolution neural network model, the simulation results of cyclic neural network model are the best with the training data set R value reaching 0.92, the verification data set The R value reaching 0.92, and the average error of the sound intensity level reaching 3.4 decibels. The modeling analysis of mobile noise monitoring data shows that the R values of the training and verification data sets are 0.96 and 0.91 respectively, and the average error is 4.3 decibels. The above two noise big data models are used to simulate regional environmental noise and road traffic noise respectively, and the simulation results are amplified to the whole urban area by using the space-time data filling technology, and an urban environmental noise simulation map with a spatial accuracy of 200 meters*200 meters and a time accuracy of minutes is established. The simulation results show that during June 2018, there were 4.24 square kilometers of noise intensity level exceeding the standard at different times in Haidian District, mainly concentrated around Wudaokou and Zhongguancun, and the time was concentrated in commuting periods such as 9:00, 18:00-19:00, and the corresponding areas of excessive noise were 0.64 square kilometers, 0.54 square kilometers and 0.58 square kilometers respectively.