神经网络用于近海水质预测的研究

运用神经网络,提出一种完全依据环境监测数据的近海水质预测模型.首先,根据以往的研究成果,确定预测模型的输入和输出因子;然后,针对训练样本序列短、群体小的特点,采用自动正则化技术避免了网络的过拟合问题;在此基础上,研究确定网络的最小结构并作适当放大,保证网络充分拟合;最后,对入海河流监测数据进行处理,实现输入和输出因子的频率一致性.经过网络训练,预测平均误差为26.46%,满足环境管理的精度要求.应用表明,这预测方法避免了机理性研究对众多基础数据的要求,原理简单,实用性强,能够为环境管理提供决策支持.

Application of neural network to prediction of coastal water quality

Through the application of NN (neural network) , a new prediction model for coastal water quality based on environmental monitoring data is proposed. Firstly, the input and output factors of the NN are determinated. Secondly , automatic regulation technology is adopted to avoid the over-fitting problem of the NN model in the thinking that the basic data serial is short. Thirdly, the smallest structure of the NN model is decided and then magnified properly to assure the model could be adequately fitted. Lastly, the data about the rivers are treated so that the input and output data have the same frequency. Then the NN model is trained and the average prediction error is 26.46%, which reaches the demand of environmental management. Through application it could be found that the extreme demand on basic data in mechanism studies could be avoided, which made the method simple, practicable and could be the decision support for environmental management.