APPLICATION OF UNCERTAINTY ANALYSIS OF ARTIFICIAL NEURAL NETWORKSFOR PREDICTING COAGULANT AND ALKALIZER DOSAGES IN A WATER TREATMENT PROCESS

ABSTRACT Artificial neural networks (ANNs) were built to predict coagulant (Model I) and alkalizer (Model II) dosages given raw and treated water parameters from a water clarifying process. Different ANN architectures were tested and optimal results were obtained with [10-10-10-01] and [08-12-12-01] nodes of input, hidden and output layers for Models I and II, respectively. Two algorithms based on GUM-S1weredevelopedto evaluate the artificial neural network parameter uncertainty and the coverage interval of model outputs. The results show that these algorithms can provide a better set of parameters for the ANN compared with the traditional training method. The present research provides a unique unifying view that considers neural networks and uncertainty analysis in a well-documented industrial case study.

摘要 本研究构建了人工神经网络（Artificial Neural Networks，ANNs），以基于净水工艺的原水与处理后水参数，预测混凝剂（模型I）与碱化剂（模型II）的投加量。研究测试了多种人工神经网络架构，其中模型I与模型II分别采用[10-10-10-01]、[08-12-12-01]的输入层、隐藏层与输出层节点配置，获得了最优结果。本研究开发了两种基于GUM-S1的算法，用于评估人工神经网络的参数不确定性与模型输出的覆盖区间。结果表明，相较于传统训练方法，此类算法可为人工神经网络提供更优的参数集合。本研究针对一项记录详实的工业案例研究，提出了兼顾神经网络与不确定性分析的独特统一视角。