Faculty of Technology and Metallurgy, University of Belgrade, Karnegijeva 4, Belgrade, 11120, Serbia.
Innovation Center of the Faculty of Technology and Metallurgy, University of Belgrade, Karnegijeva 4, Belgrade, 11120, Serbia.
Environ Sci Pollut Res Int. 2018 Apr;25(10):9360-9370. doi: 10.1007/s11356-018-1246-5. Epub 2018 Jan 18.
This paper presents an application of experimental design for the optimization of artificial neural network (ANN) for the prediction of dissolved oxygen (DO) content in the Danube River. The aim of this research was to obtain a more reliable ANN model that uses fewer monitoring records, by simultaneous optimization of the following model parameters: number of monitoring sites, number of historical monitoring data (expressed in years), and number of input water quality parameters used. Box-Behnken three-factor at three levels experimental design was applied for simultaneous spatial, temporal, and input variables optimization of the ANN model. The prediction of DO was performed using a feed-forward back-propagation neural network (BPNN), while the selection of most important inputs was done off-model using multi-filter approach that combines a chi-square ranking in the first step with a correlation-based elimination in the second step. The contour plots of absolute and relative error response surfaces were utilized to determine the optimal values of design factors. From the contour plots, two BPNN models that cover entire Danube flow through Serbia are proposed: an upstream model (BPNN-UP) that covers 8 monitoring sites prior to Belgrade and uses 12 inputs measured in the 7-year period and a downstream model (BPNN-DOWN) which covers 9 monitoring sites and uses 11 input parameters measured in the 6-year period. The main difference between the two models is that BPNN-UP utilizes inputs such as BOD, P, and PO, which is in accordance with the fact that this model covers northern part of Serbia (Vojvodina Autonomous Province) which is well-known for agricultural production and extensive use of fertilizers. Both models have shown very good agreement between measured and predicted DO (with R ≥ 0.86) and demonstrated that they can effectively forecast DO content in the Danube River.
本文应用实验设计方法对人工神经网络（ANN）进行优化，以预测多瑙河的溶解氧（DO）含量。本研究旨在获得更可靠的 ANN 模型，该模型使用较少的监测记录，同时优化以下模型参数：监测站点数量、历史监测数据数量（以年表示）以及使用的输入水质参数数量。采用 Box-Behnken 三因子三水平实验设计对 ANN 模型的时空和输入变量进行同步优化。使用前馈反向传播神经网络（BPNN）进行 DO 预测，而使用多滤波器方法（第一步采用卡方排序，第二步采用相关消除）在模型外选择最重要的输入。绝对误差和相对误差响应曲面的等高线图用于确定设计因素的最优值。从等高线图中，提出了覆盖塞尔维亚境内整个多瑙河流量的两个 BPNN 模型：一个上游模型（BPNN-UP），覆盖贝尔格莱德之前的 8 个监测站点，使用在 7 年期间测量的 12 个输入，以及一个下游模型（BPNN-DOWN），覆盖 9 个监测站点，使用在 6 年期间测量的 11 个输入参数。两个模型的主要区别在于，BPNN-UP 使用 BOD、P 和 PO 等输入，这与该模型覆盖塞尔维亚北部（伏伊伏丁那自治省）的事实是一致的，该地区以农业生产和广泛使用化肥而闻名。两个模型均显示出测量值与预测值 DO 之间非常好的一致性（R≥0.86），表明它们可以有效地预测多瑙河的 DO 含量。
