Civil and Environmental Engineering publications

Permanent URI for this collection

https://hdl.handle.net/2440/908

Browse

Browsing Civil and Environmental Engineering publications by Author "14th Water Distribution Systems Analysis Conference (24 Sep 2012 - 27 Sep 2012 : Adelaide, S.A.)"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    ItemMetadata only
    ANN forecasting of water quality in the chloraminated goldfield and agricultural pipeline system: constructive input selection
    (Engineers Australia, 2012) Wu, W.; Maier, H.; Dandy, G.; 14th Water Distribution Systems Analysis Conference (24 Sep 2012 - 27 Sep 2012 : Adelaide, S.A.)
    This paper presents an innovative method for selecting appropriate inputs for developing artificial neural network (ANN) models for water quality modeling in water distribution systems (WDSs). Input selection is an important step in the ANN model development process, particularly for complex systems with a large number of inputs. The input selection method presented in this paper is a two-step model-based constructive method using partial mutual information (PMI). In the first step, potential inputs are ordered based on their relative importance determined by their PMI values; then a model-based constructive process is used to select the most suitable inputs for the model under consideration. This algorithm significantly increases the efficiency of input selection for problems with a large number of potential inputs. The proposed method is applied to the Goldfield and Agriculture Water System (GAWS) east of Perth, Western Australia. Since the travel time in the system is unknown, four models with forecasting horizons varying from one to 10 days are developed. For all four models, the proposed input selection method effectively reduces the number of inputs from over 1600 to under 20. The models developed using the inputs selected using the proposed algorithm are able to forecast total chlorine in the system with an acceptable level of accuracy (e.g. R2 ranging from 0.71 to 0.91). More importantly, as the forecasting horizon increases, the inputs selected consist of more variables from the upstream end of the system, including the control variable, which indicates the model with a longer forecasting horizon (e.g. 10 days) better reflects the real underlying physical/chemical processes in the system.