With the high frequency automatic monitoring of surface water quality, a technique for early warning of water pollution incidents was developed using the water quality soft measurement and abnormal detection of time series. This technique takes the assumption that water pollution incidents would cause the change of typical automatic monitoring water quality parameters, and then establishes the linear relationship between the water quality parameters and online high frequency monitoring water quality parameters. Using the artificial neural network, the change of water quality parameters in a short duration was predicted; using the time series of residual error, the threshold of abnormal change was determined. Finally, early warning of pollution incidents could be achieved through detecting abnormal change based on sequential leader clustering algorithm. To verify the technique, this study takes the online monitoring data obtained from the Potomac River in Virginia, USA as a case study. The analysis of the receiver operating characteristic curve (ROC) shows that the detection accuracies of double and triple abnormal levels can reach 62.7% and 92.5%, respectively. Because the concentration level of a water pollution incident is usually significantly higher than 3times, this technique can provide a relative high accurate early warning. Compared with traditional abnormal detection methods, this technique can shorten the detection time. Along with increasing improvement of automatic monitoring facilities, this study provided a new avenue for early warning of, and prompt response to, pollution incidents.%在高频水质自动监测背景下,建立了基于软测量和水质时间序列异常检测的水体突发污染预警预报技术.假定突发污染事故会引起典型自动监测水质参数变化,采用回归分析建立水质参数和在线高频监测水质参数间的线性关系进行软测量,采用人工神经网络预测短程水质变化,建立基于预测残差的异常判断最小阈值,最终通过有序监督聚类进行水质突变检测从而对突发污染事故进行预警.采用美国弗吉尼亚州的Potomac River流域在线监测数据进行算法验证和案例分析.分析受试者工作曲线(ROC)表明:该方法对2倍异常和3倍异常水平的检测准确率分别为62.7%和92.5%,且随着异常水平的增加准确率增加,通常突发污染事故中特定污染物浓度水平一般明显高于3倍,该方法具有较高的准确率.较其他突发污染水质预警技术,该技术有效缩短了平均检测时间,为流域污染预警预报和快速应急响应提供了新途径.
展开▼
