Please use this identifier to cite or link to this item:
|Title:||Forecasting chlorine residuals in a water distribution system using an artificial neural network|
|Citation:||Regional Conference Proceedings, Adelaide, Aug. 6 2003 / Australian Water Association, South Australian Branch: pp.1-8|
|Publisher:||Australian Water Association|
|Conference Name:||Australian Water Association. South Australian Branch. Regional Conference (2003 : Adelaide, S. Aust.)|
|John Nixon, Gavin Bowden, Graeme Dandy, Holger Maier, Mike Holmes|
|Abstract:||Chlorine residual in a water distribution system (WDS) was forecast using two statistical models: multiple linear regression (MLR); and an artificial neural network (ANN). The case study was the trunk main from the Myponga water treatment plant (WTP) south of Adelaide, South Australia. The models were constructed using inputs including residual, water temperature, flow rate, turbidity, and pH. These values were obtained from telemetry and other WTP and WDS operational data, and from on-line measurements. Models were produced to forecast—over an 11 week period in Autumn 2002—residual at Almond Grove, which was calculated to be 24 hrs travel time—on average over this period—from Cactus Canyon. Accurate ANN models were produced to predict residual both 24 and 72 hrs in advance, using 1 hrly data points. Rejection of insignificant inputs, carried out using both the MLR and ANN models themselves, determined that residual from Cactus Canyon—at the time of prediction, and from Almond Grove—at 24 hrs before prediction, were the most significant. For MLR, residual from Cactus Canyon 24 hrs and 27 hrs before prediction at Almond Grove were the next two statistically significant inputs. For the ANN these inputs at Cactus Canyon, other time-lagged inputs from there, and other time-lagged inputs from Almond Grove were also significant. Model validation compared predicted against actual residual, with an independent data set not used in model creation. Residual at Almond Grove was predicted 24 hrs in advance to an accuracy comparable with that of measurement error in the field, by both the MLR and ANN models. The ANN was shown to exploit non-linear relationships in the data not able to be taken into consideration by MLR. Accurate 72 hr forecasts could be also be achieved using the ANN.|
|Keywords:||water distribution system|
artificial neural network
|Appears in Collections:||Aurora harvest|
Civil and Environmental Engineering publications
Environment Institute publications
Files in This Item:
There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.