基于多源数据机器学习的区域水质预测方法研究

Multi-source data machine learning-based study on method for regional water quality prediction

随着社会经济快速发展和水资源系统复杂性的日益增强,我国水环境质量的演变逐渐呈现跨区域、多因素耦合影响的特点。围绕大空间范围的水质预测问题,针对传统水质预测方法中对水文、气象及社会经济多因素考虑的不足,以广东省31个水质监测站在2008年到2016年间每周的水质等级数据为训练样本,选取降雨、蒸发和气温等气象指标以及GDP、总人口数、人口密度等社会经济指标为预测参数,运用支持向量机、决策树以及人工神经网络等机器学习技术,建立区域水质等级的预测模型。结果表明,机器学习方法可融合气象和社会经济等多源的、不同时空尺度的数据,对水质等级进行预测。其中,基于随机森林的预测模型表现性能最佳,预测准确率达到77.11%;基于支持向量机的预测模型次之,预测准确率达到74.99%。与现有的水质预测方法相比,该方法的计算速度快、不需要提取数据的统计特征、操作简单、能够分析社会经济因素对水质的影响,更容易在水环境治理中使用。

With the rapid socio-economic development and the progressively increase of the complexity of water resources system,the evolution of the water environment in China gradually exhibits the characteristics of the influences from cross-regional and multi-factor coupling.Focusing on the water quality prediction within a large spatial range and aiming at the insufficient considerations of hydrological,meteorological and socio-economic factors in the conventional water quality prediction methods,the weekly water quality grade data from 2008 to 2016 obtained from 31 monitoring stations in Guangdong Province are taken as the training samples and the meteorological indexes such as rainfall,evaporation and temperature as well as the socio-economic indexes,i.e.GDP,total population,population density,etc.are selected as the prediction parameters.A regional water quality grade prediction model is established by means of the machine learning techniques such as support vector machine,decision tree,artificial neural network,etc.The study results show that the machine learning methods can integrate the multi-source data of meteorology,socio-economy,etc.at different spatio-temporal scales for the prediction of water quality in which the performance of the random forest-based prediction model is the best with the accurate rate of 77.11%,while the support vector machine-based prediction is the secondary with the accurate rate of 74.99%.Compared with the existing water quality prediction methods,the method proposed herein has a faster calculation speed without the necessity of extracting the statistical features of the data,and then can be simply operated and applied to the analysis on the influences of the relevant socio-economic factors,thus can be more easily applied to the water environment improvement.