Keywords

water network, SCADA system, Linear prediction, condition-monitoring

Start Date

27-6-2018 10:40 AM

End Date

27-6-2018 12:00 PM

Abstract

From the analysis of the data captured in real time through a SCADA system, a contribution to improving the management of drinking water distribution and the early detection of anomalies is presented. In a real water network, the SCADA system must periodically acquire, store and validate the data collected by sensor measurements to achieve accurate network monitoring. For each sensor measurement, the raw data is usually represented by one-dimensional time series which must be validated before further use to ensure the reliability of the results obtained. In the present approach, we use linear predictors to verify data, detect outliers and restore missing values as well as to forecast different variables at different time intervals. The comparison of the predictions with measurements also serves to generate an error which is reported to an expert through warnings when it is unusually high. This human operator tries to associate significant prediction errors with pump configuration changes or system failures. The work mainly focuses on the predictor’s configuration at different temporal levels as well as the rapid training strategy used for an ANN (Artificial Neural Networks).

Stream and Session

B3: Sixth Session on Data Mining as a Tool for Environmental Scientists (S-DMTES-2018)

Share

COinS
 
Jun 27th, 10:40 AM Jun 27th, 12:00 PM

Linear prediction techniques for performance enhancement and maintenance of water networks using SCADA data

From the analysis of the data captured in real time through a SCADA system, a contribution to improving the management of drinking water distribution and the early detection of anomalies is presented. In a real water network, the SCADA system must periodically acquire, store and validate the data collected by sensor measurements to achieve accurate network monitoring. For each sensor measurement, the raw data is usually represented by one-dimensional time series which must be validated before further use to ensure the reliability of the results obtained. In the present approach, we use linear predictors to verify data, detect outliers and restore missing values as well as to forecast different variables at different time intervals. The comparison of the predictions with measurements also serves to generate an error which is reported to an expert through warnings when it is unusually high. This human operator tries to associate significant prediction errors with pump configuration changes or system failures. The work mainly focuses on the predictor’s configuration at different temporal levels as well as the rapid training strategy used for an ANN (Artificial Neural Networks).