神经网络模型用于数值水质模型逼近的适用性及非敏感参数的欺骗效应

Neural networks for approximating numerical water quality models:Applicability and deceptive effects of insensitive parameters

水质模型被广泛应用于水环境管理和决策,但却面临着计算时间和模型应用效率等多方面的问题;利用函数映射和逼近等方法来建立水质模型的输入-输出响应关系,可有效减少计算成本并显著改善模型效率.水质模型的输入-输出响应函数关系有多种形式,本文以其中的2种为例,并分别基于2个水质模型(零维总磷模型、WASP/EUTRO5)的案例,分析和验证了神经网络模型在响应关系逼近中的适用性.案例的结果表明:神经网络函数可以有效地用于水质模型输入-输出响应关系的逼近;当网络规模超出阈值大小时,神经网络函数逼近的准确度和泛化度对网络规模不敏感.在案例研究的基础上,推导和讨论了在神经网络模型函数映射过程中所可能出现的非敏感参数的欺骗效应,以及可能由此导致的过度预测或过低预测问题;并建议在神经网络函数逼近中,应只包含水质模型的敏感参数,以防止降低神经网络模型的准确度.

Water quality modeling （WQM） has long been used to support decision making for environmental management; however,in many cases,a water quality model requires significant computational time,which poses a limitation on using it to evaluate a large number of source management scenarios,analyze parameter uncertainty,and conduct inverse parameter estimation modeling. The efficiency of applying water quality models can be significantly improved by reducing the computational time through a functional mapping technique. A neural network （NN） based functional approximator was developed in this study to map the input-output response relationships of numerical water quality models. The applicability of the developed functional approximator is demonstrated through numerical examples of a total phosphorus （TP） model and WASP/EUTRO5. Particularly,the paper presents an analysis on the deceptive effect of insensitive model parameters in an NN functional mapping process. The study shows that a properly developed NN functional approximator can accurately approximate the input-output response relationship of a water quality model,and it is desired that only the sensitive parameters of a water quality models are included in the NN functional approximators to avoid degrading the NN model accuracy.