摘要
随着经济社会的高速发展和工业化建设程度不断提高,水环境问题已经严重影响甚至威胁了人类的健康。近年来,国家大力推行水环境的预测预警,许多专家学者利用人工神经网络等智能方法在富营养化评价及水华预测中得到了较为广泛的运用,也取得了一定成效。然而,人工神经网络的性能受到样本训练算法等方面的影响,在选取合适的神经网络模型、算法以及设置参数麻烦、耗时。随着问题复杂程度的增加,单个网络的隐层节点数将增加很多,训练时间将大大增加,从而造成训练困难。且由于训练过度或不够,往往导致泛化能力较差。为解决此问题,本文在对湖库水华形成机理深入分析的基础上,建立了BP网络的水华预测模型,并利用Bootstrap采样技术获取不同的数据集,分别训练多个BP网络,最终将多个网络进行集成用于建立太湖流域水华预测模型。通过基于Bagging算法的集成学习,可以对样本包含的信息进行充分挖掘,更全面的刻画因素之间的相互联系和变化规律。实验表明基于Bagging算法的BP网络集成模型预测结果与单个BP网络模型预测结果对比,具有较高的预测能力,从而获得了相对理想的预测效果。
With the rapid development of the economic society as well as the increasing improvement of the industrialized construction, the issue of water environment has seriously affected the health of our human beings. In recent years, forecasting and early warning towards the water environment has been vigorously implemented by our country while many experts have made use of intelligent method such as artificial neural network to make it widely applied in the field of eutrophication evaluation and prediction blooms which has achieved some results. However, the performance of artificial neural network is influenced by the algorithm of sample training so that in the process of the selection of suitable neural network model, algorithm and setting the parameters, it may take some time and effort. With the increase of complexity, the number of hidden layer nodes in a single network will increase as well as the training time which will pose a difficulty in training. Furthermore, due to the excessive or insufficient training, it may lead to a lower generalization. Through deep analysis into the formation mechanism of bloom in lakes and reservoirs, a water bloom prediction model is established based on BP artificial neural network. Bootstrap sampling method is used to get different datasets, with which to train multiple BP neural networks, that are integrated to set up a model for predicting water bloom in Taihu River Basin. Ensemble learning based on Bagging algorithm can adequately explore the information samples contain and fully reflect connections and change regularity among various factors. Experiment shows that result from Bagging-based neural network ensemble model has better predictive ability compared with single BP artificial neural network model, thus achieving ideal prediction effect.
出处
《计算机与应用化学》
CAS
CSCD
北大核心
2014年第2期140-144,共5页
Computers and Applied Chemistry
基金
国家自然科学基金资助项目(51179002)
北京市科技新星计划项目(2010B007)
北京市教委专项课题(PXM2013_014213_000044)
北京市教委科技计划面上项目(KM201110011006)
