Three-layered Feedforward artificial neural network with dropout for short-term prediction of class-differentiated Chl-a based on weekly water-quality observations in a eutrophic agricultural reservoir
Three-layered Feedforward artificial neural network with dropout for short-term prediction of class-differentiated Chl-a based on weekly water-quality observations in a eutrophic agricultural reservoir
复制标题
DOI：
10.1007/s10333-021-00874-3
复制
发表时间：
2021-10
影响因子：
2.2
通讯作者：
Ren Yamamoto;M. Harada;Kazuaki Hiramatsu;T. Tabata
中科院分区：
文献类型：
--
作者：
Ren Yamamoto;M. Harada;Kazuaki Hiramatsu;T. Tabata
为了有效管理富营养化水库，对藻类叶绿素 a (Chl-a) 进行了短期预测。该研究采用三层前馈人工神经网络，使用七年来（2012-2018）从 5 月到 11 月每周观察收集的离散水环境数据集。该网络采用监督学习构建，将某一观察日的可用数据集设置为输入变量，以确定一周后存在的总叶绿素、绿藻叶绿素和蓝藻叶绿素。从简化网络复杂度、抑制过拟合的角度出发，通过识别与绿藻和蓝藻季节变化相关的重要变量，消除水质参数的重复表达，精心选择输入变量。然而，发现网络规模缩小不足以抑制过度拟合。为了提高预测精度，引入了 dropout，即在学习过程中随机停用输入层和隐藏层中的一些节点。分析结果表明，可以实现对总叶绿素和绿藻叶绿素的足够的短期预测。使用尽可能接近所需预测日的气象数据可以克服蓝藻叶绿素a的预测准确性不足的问题。因此，该模型可以作为富营养化水库管理的有用工具，因为可以实现对优势浮游植物的短期预测，并可以相应地规划必要的缓解措施。
To effectively manage a eutrophic reservoir, short-term predictions of algae class-differentiated chlorophyll a (Chl-a) were conducted. The study adopted a three-layered feedforward artificial neural network using discrete water environment datasets collected through weekly observations from May to November over seven years (2012–2018). This network was constructed using supervised learning, and the available datasets of a certain observation day were set as the input variables to determine the total Chl-a, Chlorophyceae Chl-a, and cyanobacteria Chl-a that would exist after one week. From the viewpoint of the simplification of the network’s complexity to suppress overfitting, input variables were carefully selected by identifying the important variables related to the seasonal changes in Chlorophyceae and cyanobacteria and eliminating the duplicated expressions of water-quality parameters. However, network downsizing was found insufficient to suppress overfitting. To improve prediction accuracy, dropout was introduced, which stochastically deactivated some nodes in the input and hidden layers in the learning process. The analysis results showed that sufficient short-term predictions of total Chl-a and Chlorophyceae Chl-a may be achieved. The insufficient prediction accuracy of cyanobacteria Chl-a may be overcome using meteorological data as close as possible to the desired prediction day. Consequently, this model may serve as a useful tool for the management of eutrophic reservoirs because short-term predictions of the dominant phytoplankton can be achieved, and the necessary mitigation measures may be accordingly planned.