University of Belgrade, Faculty of Technology and Metallurgy, Karnegijeva 4, 11120 Belgrade, Serbia.
University of Belgrade, Innovation Center of the Faculty of Technology and Metallurgy, Karnegijeva 4, 11120 Belgrade, Serbia.
Sci Total Environ. 2018 Jan 1;610-611:1038-1046. doi: 10.1016/j.scitotenv.2017.08.192. Epub 2017 Aug 30.
Accurate prediction of water quality parameters (WQPs) is an important task in the management of water resources. Artificial neural networks (ANNs) are frequently applied for dissolved oxygen (DO) prediction, but often only their interpolation performance is checked. The aims of this research, beside interpolation, were the determination of extrapolation performance of ANN model, which was developed for the prediction of DO content in the Danube River, and the assessment of relationship between the significance of inputs and prediction error in the presence of values which were of out of the range of training. The applied ANN is a polynomial neural network (PNN) which performs embedded selection of most important inputs during learning, and provides a model in the form of linear and non-linear polynomial functions, which can then be used for a detailed analysis of the significance of inputs. Available dataset that contained 1912 monitoring records for 17 water quality parameters was split into a "regular" subset that contains normally distributed and low variability data, and an "extreme" subset that contains monitoring records with outlier values. The results revealed that the non-linear PNN model has good interpolation performance (R=0.82), but it was not robust in extrapolation (R=0.63). The analysis of extrapolation results has shown that the prediction errors are correlated with the significance of inputs. Namely, the out-of-training range values of the inputs with low importance do not affect significantly the PNN model performance, but their influence can be biased by the presence of multi-outlier monitoring records. Subsequently, linear PNN models were successfully applied to study the effect of water quality parameters on DO content. It was observed that DO level is mostly affected by temperature, pH, biological oxygen demand (BOD) and phosphorus concentration, while in extreme conditions the importance of alkalinity and bicarbonates rises over pH and BOD.
准确预测水质参数（WQPs）是水资源管理中的一项重要任务。人工神经网络（ANNs）常用于溶解氧（DO）预测，但通常仅检查其插值性能。本研究的目的除了插值外，还包括确定用于预测多瑙河 DO 含量的 ANN 模型的外推性能，以及评估在存在超出训练范围的值的情况下，输入的重要性与预测误差之间的关系。所应用的 ANN 是一种多项式神经网络（PNN），它在学习过程中执行最重要输入的嵌入式选择，并以线性和非线性多项式函数的形式提供模型，然后可以使用该模型对输入的重要性进行详细分析。可用的数据集包含 1912 条监测记录，涵盖 17 种水质参数，分为“常规”子集和“极端”子集。“常规”子集包含正态分布和低变异性数据，“极端”子集包含具有异常值的监测记录。结果表明，非线性 PNN 模型具有良好的插值性能（R=0.82），但在外推时不够稳健（R=0.63）。外推结果分析表明，预测误差与输入的重要性相关。即，输入的低重要性的超出训练范围的值不会显著影响 PNN 模型的性能，但它们的影响可能会因存在多个异常值监测记录而产生偏差。随后，线性 PNN 模型成功应用于研究水质参数对 DO 含量的影响。观察到 DO 水平主要受温度、pH 值、生化需氧量（BOD）和磷浓度的影响，而在极端条件下，碱度和碳酸氢盐的重要性会超过 pH 值和 BOD。
