程海富营养化机理的神经网络模拟及响应情景分析

揭示湖泊的富营养化发生机制、定量了解关键生源要素与藻类爆发的因果关联对有效改善湖泊水质和富营养化状况具有重要的科学与决策意义。本研究以云南省程海为例,建立了基于神经网络的响应模型,对富营养化机理进行了研究,并从富营养化核心驱动因子识别、神经网络模型构建与架构分析以及叶绿素a(Chl a)与TN、TP浓度降低的响应模拟几个方面对面临的科学问题进行探索。模拟结果表明,神经网络模型必须在适当的架构下才能产生科学合理的结果；程海的富营养化机制由一个氮(N)、磷(P)共限制的营养盐-藻类动力结构主导,但在此主导结构下拥有氮型限制的次级结构。基于神经网络模型模拟,推导出一系列基于湖体水质控制的Chl a响应的非线性函数,为程海的富营养化控制提供了快速决策支持。

Neural network modeling of the eutrophication mechanism in Lake Chenghai and corresponding scenario analysis

To understand the eutrophication mechanism and quantify the responsive relationship between key nutrients and algal blooms is of critical scientific and practical significance for effectively improving the water quality condition in Lake Chenghai. Although two broad families of modeling approaches, i.e., the data driven approaches, and the mechanistic modeling approaches, are both potentially applicable to exploring the relationships between nutrients and algal blooms, it was determined that only the former is viable in this case due to the severe data limitation excluding the development of a mechanistic water quality model for Lake Chenghai. Considering the data availability and the need for universal functional mapping capability, this study chose the Neural Network (NN) technology as a data-driven modeling platform for constructing the Lake Chenghai water quality model. In light of the potential deceptive effect of NN models caused by inclusion of insensitive parameters in the input nodes, the modeling analysis started with using a nonlinear curve-fitting and correlation analysis method to screen all the monitored physical and chemical parameters for identifying the key parameters. The process leads to the findings that among all the parameters, only TP and TN are qualified for being included in the Lake Chenghai model since they not only show very high correlations with the chlorophyll-a concentration, but also they are immune to the data-time coupling issues as experienced by other parameters such as inorganic nitrogen and phosphorus. Following the identification of the key input parameters, a series of NN models with various architectures were developed to explore the quantitative relationship between chlorophyll-a and nutrients in the lake. Through extensive evaluations, it was discovered that when the complexity of the NN model increased to such a level that the number of hidden node is equal or greater than 3, the NN models start to show the trait of memorization dominance, suggesting that they mimic the observed pattern through memorization rather than reasoning, hence resulting in a degradation of the generalization capability. A NN model with degraded generalization capability is generally undesired in real-world practice, therefore, the more complex network structures were discarded. As a result, only the two simple networks having respectively one and two hidden nodes were adopted as the final water quality models for Lake Chenghai. The reason of simultaneously applying both the network structures as the basis of further analysis was to account for the predictive uncertainty resulted from network configurations. The two NN models were then applied to conduct a number of scenarios for analyzing the eutrophication mechanisms in Lake Chenghai,. The modeling results show that the eutrophication in Lake Chenghai is controlled by a nested two-level limiting structure, where the dominant level is a nitrogen-phosphorus co-limiting structure, under which a secondary limiting structure dominated by nitrogen is identified. Based on the simulation results of the NN-based water quality models, a series of nonlinear functions relating chlorophyll-a concentration to water quality concentration control in Lake Chenghai were derived for supporting quick eutrophication control decision making in the future.