Estimation of Peat-Induced Color using Artificial Neural Networks

In upland peat catchments dissolved organic matter can result in high raw water colour. The inability of treatment works to effectively reduce high colour levels and remove these organics results in aesthetic water quality failure and high levels of potentially harmful disinfection by-products throughout the distribution system. This paper explores a new approach to the modelling of raw water colour using Artificial Neural Network (ANN) technology. Based on this technique, future levels of raw water colour may be forecast such that remediation can be better planned. Daily hydrometeorological data (DHMD), which includes temperature, precipitation and raw water colour, have been collected over a twenty year period (1979-1999) in a case study catchment area. Temperature and rainfall over the previous four weeks and colour from a year previously were demonstrated to be important input variables with which to configure a model to predict current colour. In addition, time was a further factor conditioning the processes which contribute to the colouration of raw water. A Time-Lagged Recurrent Network (TLRN) and a Time-Lagged Feedforward Network (TLFN) were configured and trained to predict raw water colour at the inlet to a case study treatment works. The TLRN and TLFN inferred time dependence through training features of the networks used to learn natural input-output relationships over time using feedback mechanisms and memory kernels, respectively. The models were trained with 5 years of DHMD, and tested with one historical year in order to measure prediction performance. The results showed that a Time-Lagged Recurrent Network (TLRN) was the most effective ANN model for predicting raw water colour.